# An Thai Le

> I study how robots can plan, learn, and act reliably with limited computation and data. I am an Assistant Professor at VinUniversity in Hanoi, a Visiting Professor at TU Darmstadt, and Director of Foundation AI at VinRobotics. Contact: an@robot-learning.de.

- **Research agenda and philosophy.** My main agenda is Efficient Algorithmic Robotics, guided by the geometry of computation. I study how problem structure connects to algorithm design and hardware: what can be computed independently, reused, or coordinated.
- **Parallel planning.** Our group develops tensor search methods for GPU motion planning and model predictive control. Our work includes GTMP (IEEE RA-L 2025), MTP (TMLR 2025), MPOT (NeurIPS 2023), and MPD (IEEE T-RO 2025).
- **Efficient robot policies.** We study CPU inference, quantization, and learning from fewer demonstrations. Projects include vla.simd, FoldQuantVLA, FOCA (ICML 2026), and PAINT (CoRL 2026). Raspberry Pi results report action throughput; feedback frequency is lower.
- **Humanoid control.** At VinRobotics, I lead work on locomotion, VLA models, and onboard deployment. Our team tests compliance and locomotion on heavy humanoids. The homepage distinguishes simulation results from hardware demonstrations.
- **Numerical relativity.** In my free time, I build JAX tools to study spacetime energy conditions and rocket trajectories near rotating black holes. Papers appear in Classical and Quantum Gravity and Physical Review D.
- **Teaching and service.** I advise students and research residents. My editorial roles include IEEE RA-L and ICRA 2027. Area Chair roles include ICML 2027, ICLR 2027, CoRL 2026, and RLC 2024. NeurIPS named me a Top Reviewer in 2026.
- **Background.** I completed my PhD at TU Darmstadt with Jan Peters in 2025. My thesis is *Tensor Search Methods for Vectorizing Motion Planning*. I studied at Universität Stuttgart and worked at Bosch, HLRS, and Intel.
- **Work with us.** I welcome research collaborations, student projects, invited talks, and partnerships that test algorithms on real robots. Details: https://anindex.github.io/#join
- Name variants: An Thai Le, An T. Le, Le Thai An. ORCID: 0000-0003-0929-3316.
- Publication status is listed on the homepage. BibTeX entries are in llms-full.txt.

Markdown version of https://anindex.github.io/ for AI agents and LLM tools, generated from the HTML. Site index: https://anindex.github.io/llms.txt · Markdown: https://anindex.github.io/index.md · Full text with BibTeX: https://anindex.github.io/llms-full.txt · News feed: https://anindex.github.io/feed.xml

---

Robotics · Learning · Control  
Optimization · Geometry · Relativity

# An Thai Le

Assistant Professor, [VinUniversity](https://vinuni.edu.vn/people/le-thai-an/)  
Visiting Professor, [TU Darmstadt](https://www.ias.informatik.tu-darmstadt.de/Team/AnThaiLe) · Director of Foundation AI, [VinRobotics](https://vinrobotics.net/)

I study how robots can plan, learn, and act reliably with limited computation and data.

Our group develops parallel motion planners, [data- and compute-efficient robot foundation models](https://anindex.github.io/#scaling-down), and force-aware controllers for heavy humanoids. We study locomotion and manipulation where robots must reason about physical contact.

- <an@robot-learning.de>
- [Google Scholar](https://scholar.google.com/citations?user=k0r0tfUAAAAJ)
- [GitHub](https://github.com/anindex)
- [OpenReview](https://openreview.net/profile?id=~An_Thai_Le1)
- [ORCID](https://orcid.org/0000-0003-0929-3316)
- [LinkedIn](https://www.linkedin.com/in/anindex/)
- [X](https://x.com/an_thai_le)

[Interested in working together?](https://anindex.github.io/#join) I welcome curious students and collaborators.

## Latest news

- Nov 2026 I'm happy to share that **[AAC](https://arxiv.org/abs/2604.20744)** has been accepted by [TMLR](https://jmlr.org/tmlr/)! Many thanks to Vien for the collaboration, and to the TMLR reviewers and editors for their valuable feedback that made the paper much better! The [code](https://github.com/anindex/aac) is online.
- Oct 2026 **[CrossBFM](https://dotandung.github.io/crossbfm/)** was selected for an **oral presentation** at the [CoRL 2026 Pretrain-to-Adapt workshop](https://corl2026-robotcl.github.io/#program) on November 12! Congratulations to Tan-Dzung Do, Tuan Dat Phuong, and the whole team on this recognition!
- Oct 2026 I'm happy to serve as an **Associate Editor** for [IEEE Robotics and Automation Letters (RA-L)](https://www.ieee-ras.org/publications/ra-l) and an **Area Chair** for [ICML 2027](https://icml.cc/). I'm looking forward to working with authors and reviewers, and learning from your ideas!
- Sep 2026 Congratulations to my student Ho Trong Bao and the team! **[Start Right, Arrive Right](https://paint-action-chunking.github.io/)** (PAINT) has been accepted at [CoRL](https://www.corl.org/) 2026! It was also selected for a **spotlight** at the [RSS 2026 Diff4RL workshop](https://rss2026-diffusion-robot-learning.github.io/#spotlights). PAINT connects action chunks smoothly during asynchronous execution, without retraining.
- Sep 2026 I'm glad to contribute as an **Associate Editor** for [ICRA 2027](https://www.ieee-ras.org/conferences-workshops/fully-sponsored/icra) and an **Area Chair** for [ICLR 2027](https://iclr.cc/). Looking forward to your papers and the discussions ahead!

Earlier news

- Sep 2026 Some good news from my physics side project: **[Observer-robust energy condition verification for warp drive spacetimes](https://doi.org/10.1088/1361-6382/aeb25a)** (warpax) has been accepted by *Classical and Quantum Gravity* (published online October 2026)! I use vectorized automatic differentiation in JAX to check energy conditions in warp drive spacetimes. Many thanks to the reviewers for helping me improve the paper! [Code](https://github.com/anindex/warpax) is online.
- Sep 2026 Glad to share that **[Training Non-Differentiable Networks via Optimal Transport](https://openreview.net/forum?id=8mlcqTTMuU)** (PolyStep) has been accepted by [TMLR](https://jmlr.org/tmlr/)! PolyStep trains non-differentiable networks through optimal transport with only forward evaluations. Thank you to the reviewers and editors for their thoughtful feedback! [Code](https://github.com/anindex/polystep) is online.
- Jul 2026 It's been a busy few months with our [VinUniversity](https://vinuni.edu.vn/) students and [VinRobotics](https://vinrobotics.net/) residents. We have new preprints on VLA fine-tuning, equivariance, noise selection, and inference, plus payload-robust locomotion and whole-body compliance. I'm lucky to work with this team! Thank you all for the care and effort behind these papers! The [papers](https://anindex.github.io/#publications) are below.
- Jul 2026 **[StructSAM](https://arxiv.org/abs/2603.07307)**, token merging for Segment Anything, was presented at the ICML 2026 AdaptFM workshop. Congratulations to the team!
- Jun 2026 I'm happy to serve as an **Area Chair** for [CoRL 2026](https://www.corl.org/). Looking forward to reading your work and exchanging ideas with the community!
- Jun 2026 Thank you to the organizers for inviting me to speak about *Designing Planning Algorithms for the Era of Parallelism* at the **[ICRA 2026 Workshop on Robot Architecture](https://sites.google.com/view/roboarch-icra26/home)** and the **[2nd Next-Gen Robot Learning Symposium](https://sites.google.com/view/robot-learning-symposium)**. The [talk slides](https://anindex.github.io/slides/roboarch/) are online!
- Jun 2026 Happy to share **[CLOT](https://mengguo.github.io/personal_site/papers/pdf/zhang2026clot.pdf)**, our work on multi-robot planning with optimal transport and signal temporal logic, at [ICRA 2026](https://2026.ieee-icra.org/) in Vienna. Many thanks to [Meng Guo](https://mengguo.github.io/personal_site/)'s group at Peking University for the collaboration!
- May 2026 I'm delighted that **[Rarity of rocket-driven Penrose extraction in Kerr spacetime](https://doi.org/10.1103/f3jn-7wcp)** has been accepted by [Physical Review D](https://journals.aps.org/prd/) (published June 2026, Phys. Rev. D 113, 124036)! Thank you to the reviewers and editors for their helpful feedback!
- Apr 2026 **[FOCA: Future-Oriented Conditioning for Data-Efficient Vision-Language-Action Adaptation](https://focavla.github.io/)** has been accepted at [ICML 2026](https://icml.cc/virtual/2026/poster/66754)! Congratulations to co-first authors Duc Minh Nguyen, Nghiem Tuong Diep and Binh Gia Nguyen, and thanks to the whole team!
- Feb 2026 **[DoublyAware](https://doi.org/10.1109/LRA.2025.3648611)**, uncertainty-aware TD-MPC for humanoid locomotion, is now published in [IEEE RA-L](https://www.ieee-ras.org/publications/ra-l) (vol. 11, no. 2). Congratulations, Khang Nguyen! It's been a pleasure working on this together!
- Oct 2025 I joined [VinUniversity](https://vinuni.edu.vn/) as **Assistant Professor** and [TU Darmstadt](https://www.tu-darmstadt.de/) as **Visiting Professor**. I'm looking forward to working with the students and colleagues at both places!
- Aug 2025 I joined **[VinRobotics](https://vinrobotics.net/)** as **Director of Foundation AI**, working on RL locomotion, humanoid VLA architectures, and onboard deployment. I'm excited to work with the team and put these algorithms to use on real hardware!
- Aug 2025 I defended my PhD thesis on *Tensor Search Methods for Vectorizing Motion Planning* at [TU Darmstadt](https://www.tu-darmstadt.de/)! The [full thesis](https://tuprints.ulb.tu-darmstadt.de/entities/publication/41c837ff-8bab-4ad1-acaf-4d4602f41b26) is on TUprints. A heartfelt thank you to [Prof. Jan Peters](https://www.ias.informatik.tu-darmstadt.de/Team/JanPeters) for tolerating my mischief and supporting me throughout. I could not have asked for a better advisor! I'm also grateful to [Prof. Siddhartha Srinivasa](https://goodrobot.ai/) for serving as external examiner, and to the rest of the committee for their time and thoughtful feedback!
- Jul 2025 **[Model Tensor Planning](https://openreview.net/forum?id=fk1ZZdXCE3)** has been accepted at [TMLR](https://jmlr.org/tmlr/) 2025 and [ICLR](https://iclr.cc/) 2026 (J2C track)! Many thanks to my co-authors for the collaboration and to the reviewers for feedback that made the paper better!
- Jul 2025 **[Global Tensor Motion Planning](https://sites.google.com/view/gtmp/home)** is now published in [IEEE RA-L](https://www.ieee-ras.org/publications/ra-l) 2025 and accepted at [ICRA](https://www.ieee-ras.org/conferences-workshops/fully-sponsored/icra) 2026! Thanks to my co-authors and the reviewers for helping bring this work together!
- Jun 2025 **[Motion Planning Diffusion](https://sites.google.com/view/motionplanningdiffusion)** is now published in [IEEE T-RO](https://www.ieee-ras.org/publications/t-ro) 2025! Thanks to João and all our co-authors for the collaboration, and to the reviewers for their helpful feedback!
- May 2025 Thank you to my hosts at **[RMIT University](https://www.rmit.edu.vn/)**, **[Rice University](https://www.rice.edu/)**, and **[HUST](https://hust.edu.vn/)** for inviting me to discuss tensor search methods for motion planning!

## Research

### Efficient Algorithmic Robotics

My research agenda is to make capable robot behavior possible within practical limits on computation, memory, and data. I approach efficiency through algorithm design: how we represent a problem, organize its computation, and connect the result to the robot's physical behavior.

The philosophy behind this work is the [geometry of computation](https://anindex.github.io/#geometry-of-computation). I look for representations that reveal which operations can run in parallel, which experiences can be shared, and which physical constraints must stay coupled. This connects our work in motion planning, robot foundation models, and humanoid control.

01 / Problem representation

**[Plan in parallel](https://anindex.github.io/#parallel-planning)** Arrange candidate paths so a GPU can evaluate them together. [GTMP · multipartite search graph](https://arxiv.org/abs/2411.19393)

02 / Robot foundation models

**[Fit models to robot hardware](https://anindex.github.io/#scaling-down)** Reduce memory and computation while measuring what the robot can still do. [vla.simd · native SIMD computation](https://arxiv.org/html/2609.24274v1#S4.F2)

03 / Whole-body control

**[Respond to physical contact](https://anindex.github.io/#contact-control)** Coordinate the whole body as contact forces and support change. [CompliantWBC · force response](https://compliantwbc.github.io/#demos)

01 / Planning & optimization

### Representations for efficient algorithms

How can a change of representation simplify an algorithm and expose more parallel computation?

Many planners make a sequence of small decisions: sample a configuration, check an edge, and decide where to search next. That structure can limit how much work runs in parallel. In tensor motion planning, we arrange candidate configurations in layers so a GPU can evaluate many possible paths together. Changing the representation changes the operations available to the planner.

Tensor planning, optimal transport, and learned trajectory priors offer different ways to organize this search. I study how these choices affect solution quality, path diversity, and the value of additional computation. Our completeness and convergence results make the conditions for these guarantees explicit. Runtime and memory measurements then help us judge whether it fits a robot's budget.

[GTMP](https://anindex.github.io/#pub-gtmp) · [Anytime GTMP](https://anindex.github.io/#pub-anytime-gtmp) · [MPOT](https://anindex.github.io/#pub-mpot) · [MTP](https://anindex.github.io/#pub-mtp) · [MPD](https://anindex.github.io/#pub-mpd-tro) · [CLOT](https://anindex.github.io/#pub-clot) · [PolyStep](https://anindex.github.io/#pub-polystep) · [AAC](https://anindex.github.io/#pub-aac)

[GTMP](https://anindex.github.io/#pub-gtmp) · parallel motion planning. [Layered search graph](https://anindex.github.io/images/gtmp.png)

[MTP](https://anindex.github.io/#pub-mtp) · tensor search for model predictive control.

02 / Robot foundation models

### Scaling down robot foundation models

How much capability can we retain with less data, memory, and compute?

I want vision-language-action (VLA) models to be useful on hardware a robot can carry. This means studying both what a policy learns and how its computations run. We remove redundant layers and use low-bit arithmetic to reduce inference cost. Geometric symmetries and supervision from future observations help policies learn from fewer demonstrations.

A smaller model still needs an implementation that suits the hardware. Memory layout, data movement, and shared operations can matter as much as parameter count. We also use search and retrieval to guide frozen policies during execution. The test is whether the robot retains useful behavior: we measure task success alongside response time, memory use, and data requirements.

#### Raspberry Pi 5 · CPU-only inference

With vla.simd, IMPACT supplies **33.5 actions/s in FP32** and **81.2 in INT8** after 90 seconds of sustained operation. These rates measure action throughput. Feedback frequency is lower. [Paper results](https://arxiv.org/html/2609.24274v1#S6.SS2)

[FoldQuantVLA](https://anindex.github.io/#pub-foldquantvla) · [vla.simd](https://anindex.github.io/#pub-vlasimd) · [vla.cpp](https://anindex.github.io/#pub-vlacpp) · [CLP](https://anindex.github.io/#pub-clp) · [FOCA](https://anindex.github.io/#pub-foca) · [FINE](https://anindex.github.io/#pub-fine) · [EquiVLA](https://anindex.github.io/#pub-equivla) · [PAINT](https://anindex.github.io/#pub-paint) · [SDN](https://anindex.github.io/#pub-sdn-vla) · [R2A](https://anindex.github.io/#pub-r2a)

[vla.simd / IMPACT](https://vla-simd.github.io/) · "Put the tape into the box." SO-101, original speed.

[PAINT](https://anindex.github.io/#pub-paint) · connecting action chunks without retraining. 2× speed.

03 / Humanoid control

### Whole-body control through contact

How should a humanoid use its whole body to respond to contact and changing loads?

A pull on a humanoid's arm changes how its torso and legs must move to keep balance. Contact makes these parts of the body a coupled system. I study how controllers can respond through the whole body, using force estimates from onboard sensors, parallel foothold search, and compliance training. We test these ideas on 70–85 kg humanoids under contact and changing loads.

The same question extends across robot bodies: which parts of a behavior can transfer, and which depend on the hardware? CrossBFM shares a behavior space across humanoids, with the strongest transfer between similar bodies. In [γ₀](https://anindex.github.io/#pub-gamma-zero), we train one motion policy across many robot bodies. These projects help us study how shared behavior and body-specific control can work together.

[γ₀](https://anindex.github.io/#pub-gamma-zero) · [CrossBFM](https://anindex.github.io/#pub-crossbfm) · [CompliantWBC](https://anindex.github.io/#pub-compliantwbc) · [TACT-ful](https://anindex.github.io/#pub-tactful) · [DoublyAware](https://anindex.github.io/#pub-doublyaware) · [Hardware experiments](https://anindex.github.io/#humanoids)

[CompliantWBC](https://anindex.github.io/#pub-compliantwbc) · whole-body response to a downward pull.

[CrossBFM](https://anindex.github.io/#pub-crossbfm) · shared behavior across humanoids.

Research philosophy

### The geometry of computation

I use the geometry of computation as a design philosophy: let the relationships in a problem shape its algorithms and their implementation. Nearby trajectories can share work during search. A rotated scene may call for a correspondingly rotated action. Contact couples forces across a robot's body. I want to understand what these relationships let us compute independently, reuse, or coordinate.

1.  **Problem geometry** Trajectories, symmetries, contact
2.  **Computational structure** Batches, shared work, coordination
3.  **Hardware organization** Parallel lanes, memory, native kernels

The connection becomes concrete in the implementation. [Tensor planning](https://anindex.github.io/#pub-gtmp) organizes candidate paths into regular batches that parallel hardware can evaluate. [vla.simd](https://anindex.github.io/#pub-vlasimd) arranges shared computation around CPU instructions and memory access. [Foothold search](https://anindex.github.io/#pub-tactful) brings parallel candidate evaluation into locomotion. Choosing a representation is already an algorithmic decision: it changes which operations become cheap and which dependencies remain.

My longer-term goal is to connect these planners and policies to controllers that regulate motion and compliance. I want to test the combined system on humanoid tasks involving contact, within the robot's computation budget. Alongside task success, we need to understand how its data requirements, memory use, and response time change as tasks become harder.

A personal interest

### Computational general relativity

[warpax](https://anindex.github.io/#pub-warpax) · expansion and energy-condition diagnostics for an Alcubierre-bubble parameter sweep.

I write numerical tools for general relativity in my free time because, to me, the physics is art. [warpax](https://anindex.github.io/#pub-warpax) computes spacetime curvature with automatic differentiation in JAX. It checks energy conditions for all observers at a point through matrix inequalities. These tests alone cannot establish whether a warp drive is physically realizable.

The [Kerr-flyby study](https://anindex.github.io/#pub-penrose) simulates rocket trajectories to test when energy extraction from a rotating black hole still allows escape. I enjoy turning geometric equations into numerical experiments that I can run and inspect. As in robotics, the representation matters: I batch the computation and make the assumptions explicit.

[warpax · Classical and Quantum Gravity](https://anindex.github.io/#pub-warpax) · [Penrose extraction · Physical Review D](https://anindex.github.io/#pub-penrose)

In the news: warpax in [Popular Mechanics](https://www.popularmechanics.com/science/a71362600/scientists-say-physical-warp-drive-possible/) (May 2026); VinRobotics' visit to [DFKI](https://iml.dfki.de/news/promising-exchange-with-vinrobotics-on-future-collaboration/); VinRobotics humanoids at ICRA 2026 in [Báo Tin Tức](https://baotintuc.vn/cong-nghe-va-kien-tao/robot-hinh-nguoi-cua-viet-nam-gay-an-tuong-tai-dien-dan-hang-dau-the-gioi-20260605093328796.htm).

## Humanoids at work

Hardware experiments with our [VinRobotics](https://vinrobotics.net/) team: handling parts, performing martial arts, responding to contact, and walking under load.

[View the full kitting animation](https://anindex.github.io/images/kitting.gif)

### Kitting

A humanoid handles parts in a kitting pilot.

[View the full humanoid high-kick animation](https://anindex.github.io/images/vovinam.gif)

### Humanoid high kick

Whole-body motion imitation on a humanoid.

[View the full whole-body compliance animation](https://anindex.github.io/images/compliant_wbc.gif)

### [Compliant whole-body control](https://anindex.github.io/#pub-compliantwbc)

Coordinating the whole body to respond to contact.

[View the full TACT-ful stair descent animation](https://anindex.github.io/images/tactful_down_stair_20kg.gif)

### [Stair descent with a 20 kg payload](https://anindex.github.io/#pub-tactful)

TACT-ful combines foothold planning with compliance training.

[More humanoid videos](https://anindex.github.io/#demos) · [Research behind the demos](https://anindex.github.io/#research)

## Publications

Papers and ongoing projects, each with its current status. \* marks co-first or co-last authors. A blue line marks representative papers. Select a figure to view it at full size. See also [Google Scholar](https://scholar.google.com/citations?user=k0r0tfUAAAAJ).

### 2026

1.  #### [RoboWorM: Evaluating Movement Realism, Consistency and Acceleration Affects in Embodied Video World Models](https://arxiv.org/abs/2610.33807)

    Doanh Le Thien\*, Viet-Thanh Nguyen\*, Pham Tri Quang\*, Nhi Pham, Christopher Wewer, Minh-Loi Nguyen, Hoang H. Tran, Tho Le, Tung M. Luu, Dung D. Le, Vu N. Duong, Jonas Fischer, Ngan Le, An Thai Le, Daniel Sonntag, Gim Hee Lee, James Zou, Mathias Niepert, Jan Peters, Duy M. H. Nguyen, Ngo Anh Vien

    RoboWorM benchmarks robot video models on 1,605 episodes across single-arm, bimanual, and humanoid robots. KIVA checks kinematic stability, and RIGIS checks consistency across camera views, without vision-language judges. We compare nine acceleration methods on three models to measure the tradeoff between speed and physical fidelity.

    Under review · 2026 · [arXiv](https://arxiv.org/abs/2610.33807) · [PDF](https://arxiv.org/pdf/2610.33807) · [Project page](https://roboworm.github.io/roboworm/)

2.  #### [SeeBFM: Behavior Foundation Model That Sees](https://capfab.github.io/seebfm-media/)

    Thanh Ly, Truong-Duy Dang, Hung Nguyen, Duy M. H. Nguyen, Vien Anh Ngo, An Thai Le

    SeeBFM adds terrain perception to BFM-Zero while keeping reward, motion, and language prompts. A student model predicts a height scan from two depth cameras to guide steps over stairs, gaps, and moving obstacles. Tests are in simulation; hardware tests are planned.

    Under review · 2026 · [Project page](https://capfab.github.io/seebfm-media/) · [Playground](https://capfab.github.io/seebfm-media/playground/)

3.  #### [γ₀: A Generalist Policy for Multi-Embodiment Motion Control](https://gamma-zero.com/)

    [Nico Bohlinger](https://www.ias.informatik.tu-darmstadt.de/Team/NicoBohlinger)\*, Bo Ai\*, Dichen Li, Tongzhou Mu, …, Tan-Dzung Do, …, An Thai Le, [Yue Wang](https://yuewang.xyz/), Davide Scaramuzza, Guanya Shi, Marco Hutter, Mac Schwager, Georgia Chalvatzaki, Henrik I. Christensen, Hao Su, Chelsea Finn, [Jan Peters](https://www.ias.informatik.tu-darmstadt.de/Member/JanPeters)

    One motion policy controls humanoids and quadrupeds at the joint level, with outdoor hardware demonstrations. We vary robot bodies during training to learn what transfers between them. Try different robots in the interactive playground.

    Ongoing project · 2026 · [Project page](https://gamma-zero.com/) · [Playground](https://play.gamma-zero.com/)

4.  #### [CrossBFM: Distilling a Shared Latent Behavior Space Across Humanoid Embodiments](https://arxiv.org/abs/2609.38087)

    T.D. Do\*, T.D. Phuong\*, N. Bohlinger, C.T. Trinh, S. Ju, V.A. Ngo, J. Peters, X. Wang\*, An Thai Le\*

    CrossBFM shares a behavior space across humanoids for motion tracking, goal reaching, and reward prompts, with hardware demonstrations. We train the encoder on retargeted motion without simulation, then train a controller for each robot. Transfer is strongest between similar bodies.

    Under review · CoRL 2026 Pretrain-to-Adapt Workshop (**Oral**) · [arXiv](https://arxiv.org/abs/2609.38087) · [Project page](https://dotandung.github.io/crossbfm/) · [Workshop](https://corl2026-robotcl.github.io/#program)

5.  #### [Representative paper: CompliantWBC: Whole-Body Compliance for Heavy Humanoids via Force Latent Estimation and Residual Impedance Targets](https://compliantwbc.github.io/)

    T.D. Do\*, C.T. Trinh\*, T.D. Phuong, T.D. Dang, C. Le, T. Ly, V.A. Ngo, An Thai Le

    CompliantWBC estimates external forces from proprioception and adjusts impedance targets so a 70 kg humanoid yields to contact. The passivity guarantee is local and assumes fixed stiffness.

    Under review · 2026 · [Project page](https://compliantwbc.github.io/) · [PDF](https://compliantwbc.github.io/paper/compliant_wbc.pdf)

6.  #### [TACT-ful: Multi-Channel Terrain Affordance and Compliance Training for Payload-Robust Perceptive Humanoid Locomotion](https://arxiv.org/abs/2606.20645)

    T. Ly\*, T.D. Dang\*, C. Le, T.D. Do, T.D. Phuong, C.T. Trinh, V.A. Ngo, An Thai Le

    TACT-ful combines GPU-parallel foothold search with compliance training for uneven terrain and changing payloads. We report quantitative simulation results and qualitative hardware demonstrations.

    Under review · 2026 · [arXiv](https://arxiv.org/abs/2606.20645) · [Project page](https://fai-rl-tech.github.io/tact-locomotion.github.io/)

7.  #### [Pretraining Foundation Policies for Perceptive Humanoid Locomotion with Offline RL (SOFIE)](https://silentgradient.github.io/SOFIE/)

    SOFIE learns from offline expert rollouts and gives better footsteps more weight. One policy walks three humanoids over stepping stones, stairs, and rough ground. Simulation tests cover twelve humanoids across two simulators. Pretraining also helps adaptation to an unseen humanoid.

    Under review · 2026 · [Project page](https://silentgradient.github.io/SOFIE/)

8.  #### [Terrain Traversal Without Motion Priors: Expert-Routed Humanoid Locomotion at Deployment Fidelity](https://fai-rl-tech.github.io/moe-perceptive/)

    A terrain classifier selects among five locomotion experts, and gradient steps refine each foothold. Training uses no motion-capture retargeting, generative motion prior, or distillation. The policy completes an 85 m course in 18 of 20 runs in a second simulator configured for deployment. Hardware tests are pending.

    Under review · 2026 · [Project page](https://fai-rl-tech.github.io/moe-perceptive/)

9.  #### [DoublyAware: Dual Planning and Policy Awareness for Temporal Difference Learning in Humanoid Locomotion](https://doi.org/10.1109/LRA.2025.3648611)

    Khang Nguyen, An Thai Le, Jan Peters, Minh Nhat Vu

    DoublyAware handles planning and policy uncertainty separately in TD-MPC. Conformal filtering screens candidate trajectories, while a group-relative trust region constrains policy updates. We evaluate it on a simulated H1-2 humanoid.

    IEEE RA-L 2026 · [DOI](https://doi.org/10.1109/LRA.2025.3648611) · [arXiv](https://arxiv.org/abs/2506.12095)

10. #### Representative paper: [FoldQuantVLA: Native Low-Bit Quantization of Vision-Language-Action Models via Consistent Folding](https://arxiv.org/abs/2609.24433)

    Hung T. Ho, Khanh D. Nguyen, Quang D. Nguyen, Thanh Q. Duong, Ngan Le, Meng Guo, Vien A. Ngo, An Thai Le

    FoldQuantVLA enables native low-bit VLA inference without retraining. W4A4 projections run 1.20–1.33× faster than floating-point TensorRT on Jetson AGX Orin. Selective INT8 improves action fidelity and measured success on real robots.

    Preprint · 2026 · [arXiv](https://arxiv.org/abs/2609.24433) · [Project page](https://review-artifact-27f4.github.io/foldquantvla/) · [Code](https://github.com/cair-vinuni/FoldQuantVLA)

11. #### Representative paper: [vla.simd: Efficient CPU Inference for Language-Conditioned Manipulation](https://arxiv.org/abs/2609.24274)

    Khanh D. Nguyen, Hoang M. Truong, An Thai Le

    vla.simd runs VLA policies on CPUs with shared SIMD kernels, demonstrated on SO-101 and UR10e. On Raspberry Pi 5, IMPACT supplies 33.5 actions/s in FP32 after 90 seconds of sustained operation. Feedback frequency is lower than action throughput.

    Preprint · 2026 · [arXiv](https://arxiv.org/abs/2609.24274) · [Project page](https://vla-simd.github.io/) · [Code](https://github.com/cair-vinuni/vla.simd)

12. #### [Representative paper: Training Non-Differentiable Networks via Optimal Transport (PolyStep)](https://openreview.net/forum?id=8mlcqTTMuU)

    An Thai Le

    PolyStep trains non-differentiable networks with forward evaluations alone. Optimal transport weights candidate parameter updates. We compare gradient-free methods with matched evaluation budgets.

    TMLR 2026 · [OpenReview](https://openreview.net/forum?id=8mlcqTTMuU) · [arXiv](https://arxiv.org/abs/2605.01928) · [Code](https://github.com/anindex/polystep)

13. #### [Representative paper: Observer-robust energy condition verification for warp drive spacetimes (warpax)](https://doi.org/10.1088/1361-6382/aeb25a)

    An Thai Le

    warpax uses automatic differentiation in JAX to check energy conditions in warp drive spacetimes. It combines 4×4 matrix inequalities over all observers at a point with interval bounds for certified decisions.

    Class. Quantum Grav. 2026 · [DOI](https://doi.org/10.1088/1361-6382/aeb25a) · [arXiv](https://arxiv.org/abs/2602.18023) · [Code](https://github.com/anindex/warpax)

14. #### [Representative paper: Start Right, Arrive Right: Asynchronous Execution via Initial Noise Selection (PAINT)](https://arxiv.org/abs/2606.19774)

    T.B. Ho, Q.T. Nguyen, T.L. Ha, G.B. Nguyen, V.T. Nguyen, L. Dinh, M.N. Vu, D.M.H. Nguyen, An Thai Le, V.A. Ngo

    PAINT selects initial noise through flow inversion to connect action chunks smoothly during asynchronous execution. It needs no retraining or inference-time gradients. We evaluate it on 12 simulated benchmarks and 6 real manipulation tasks.

    CoRL 2026 · RSS 2026 Diff4RL Workshop (Spotlight) · [arXiv](https://arxiv.org/abs/2606.19774) · [Project page](https://paint-action-chunking.github.io/) · [Repository](https://github.com/cair-vinuni/paint-action-chunking) · [Workshop](https://rss2026-diffusion-robot-learning.github.io/#spotlights)

15. #### Representative paper: [FOCA: Future-Oriented Conditioning for Data-Efficient Vision-Language-Action Adaptation](https://icml.cc/virtual/2026/poster/66754)

    D.M. Nguyen\*, N.T. Diep\*, G.B. Nguyen\*, T.B. Ho, D.T. Le, Q.T. Nguyen, T.L. Ha, V.N. Tran, B. Thach, X.N. Tran, T.A. Tran, A. Habuda, P.L. Møller, T.N. Le, D. Sonntag, M. Niepert, K.D. Doan, V.N. Duong, H. Ngo, M.N. Vu, D.M.H. Nguyen\*, An Thai Le\*, V.A. Ngo\*

    FOCA learns from future interactions and goals to adapt VLA policies with few demonstrations. It reaches 95.7% success on LIBERO with 20 demonstrations per task. We also test it in RoboCasa and on real robots.

    ICML 2026 · [arXiv](https://arxiv.org/abs/2606.20867) · [Project page](https://focavla.github.io/) · [Code](https://github.com/cair-vinuni/FOCA)

16. #### [CLOT: Multi-Robot Motion Planning via Collaborative Optimal Transport under Signal Temporal Logic Tasks](https://mengguo.github.io/personal_site/papers/pdf/zhang2026clot.pdf)

    Y. Zhang, Y. Zhang, An Thai Le, M. Guo

    CLOT uses collaborative optimal transport to plan multi-robot trajectories for tasks specified in signal temporal logic. It updates trajectories without gradients, with tests in simulation and on hardware.

    ICRA 2026 · [PDF](https://mengguo.github.io/personal_site/papers/pdf/zhang2026clot.pdf)

17. #### [StructSAM: Structure- and Spectrum-Preserving Token Merging for Segment Anything Models](https://arxiv.org/abs/2603.07307)

    D.M.H. Nguyen, T.A. Tran, D. Nguyen, S. Xie, T.Q. Nguyen, M.T.N. Truong, D. Palenicek, An Thai Le, M. Barz, E. Hannus, T. Nguyen, T. Dam, T.N. Le, N. Le, M. Vu, K. Doan, V. Ngo, P. Xie, J. Zou, D. Sonntag, J. Peters, M. Niepert

    StructSAM merges redundant tokens in Segment Anything's encoder while preserving boundaries and prompt regions. We measure the tradeoff between computation and segmentation accuracy.

    ICML 2026 Workshop (AdaptFM) · [arXiv](https://arxiv.org/abs/2603.07307) · [OpenReview](https://openreview.net/forum?id=K41rNB0F8j) · [Workshop](https://icml.cc/virtual/2026/75281)

18. #### [Rarity of rocket-driven Penrose extraction in Kerr spacetime](https://journals.aps.org/prd/abstract/10.1103/f3jn-7wcp) (Penrose rocket)

    An Thai Le

    I simulate 320,000 rocket flybys around rotating black holes to test energy extraction and escape. Successful cases are rare in broad scans, concentrated at high black-hole spin and relativistic exhaust speeds.

    Phys. Rev. D 2026 · [DOI](https://doi.org/10.1103/f3jn-7wcp) · [arXiv](https://arxiv.org/abs/2601.19616) · [Code](https://github.com/anindex/penrose_process)

19. #### [Anytime Global Tensor Motion Planning (Anytime GTMP)](https://arxiv.org/abs/2608.25830)

    S. Coumar, An Thai Le\*, Z. Kingston\*

    Anytime GTMP uses black-box local planners to find more paths and reduce their cost as computation continues. Asymptotic optimality holds under stated clearance and sampling assumptions.

    Under review · 2026 · [arXiv](https://arxiv.org/abs/2608.25830) · [Code](https://github.com/CoMMALab/anytime_gtmp)

20. #### [TD-MTP: Coupling Structured Planning with Multi-Step Value Learning](https://tdmtp.github.io/)

    We test how tensor-graph planning interacts with one-step and TD(λ) value targets across ten paired tasks. On humanoid stairs, each change alone lowers return, but the combination improves on the baseline. Clear evidence of this interaction appears on two of the ten tasks.

    Under review · 2026 · [Project page](https://tdmtp.github.io/)

21. #### [EquiVLA: A General Framework for Rotationally Equivariant Vision-Language-Action Models](https://arxiv.org/abs/2606.19784)

    T.L. Ha\*, Q.T. Nguyen\*, T.B. Ho\*, L. Dinh, M.D. Nguyen, G.B. Nguyen, T.Q. Pham, M.N. Vu, D.M.H. Nguyen, An Thai Le, V.A. Ngo

    EquiVLA makes actions rotate with the input in pretrained VLA models. The perception module is approximately equivariant; the action head is exactly equivariant. We test it with GR00T N1.5 on LIBERO.

    Under review · 2026 · [arXiv](https://arxiv.org/abs/2606.19784) · [Project page](https://equivla.github.io/)

22. #### [vla.cpp: A Unified Inference Runtime for Vision-Language-Action Models](https://arxiv.org/abs/2606.08094)

    K.D. Nguyen, H.T. Ho, C.T. Nguyen, T.Q. Duong, L.D. Le, D.M.H. Nguyen, V.A. Ngo, An Thai Le

    vla.cpp runs eleven VLA models in a C++ runtime for onboard inference. In the reported test, BitVLA completes all 200 simulated LIBERO-Object episodes with inference on an 8 GB Jetson Orin Nano.

    Under review · 2026 · [arXiv](https://arxiv.org/abs/2606.08094) · [Project page](https://fai-modelopt-tech.github.io/vla-cpp.github.io/) · [Code](https://github.com/VinRobotics/vla.cpp)

23. #### [Finetuning Vision-Language-Action Models Requires Fewer Layers Than You Think (CLP)](https://arxiv.org/abs/2606.20246)

    G.B. Nguyen, T.B. Ho, T.L. Ha, K. Vo, P.L. Møller, Q.T. Nguyen, L. Dinh, T.M. Luu, T. Dam, V. Duong, T. Le, N.D.Q. Bui, M. Vu, T.N. Le, An Thai Le, N. Le, D. Sonntag, J. Zou, J. Peters, D.M.H. Nguyen, V.A. Ngo

    CLP removes redundant VLA layers before fine-tuning, using one pass to measure representation similarity. We study how model depth and training time affect task performance.

    Under review · 2026 · [arXiv](https://arxiv.org/abs/2606.20246) · [Project page](https://clpvla.github.io/) · [Code](https://github.com/cair-vinuni/CLP_VLA)

24. #### [Self-Improving VLA Policies: Selected Diffusion Noise for Spurious-Robust Action Smoothing (SDN)](https://arxiv.org/abs/2606.14084)

    D.M. Nguyen, B.N. Dao, T.M. Luu, B.G. Nguyen, V. Tong, A. Liu, V.N. Duong, D.D. Le, D. Sonntag, T. Le, N. Le, J. Peters, An Thai Le, M.N. Vu, M. Niepert, K.D. Doan, D.M.H. Nguyen, V.A. Ngo

    SDN selects initial noise for a frozen VLA. It filters for action candidates grounded in objects, then executes the smoothest candidate. Tests cover simulation and real robots, without policy retraining.

    Under review · 2026 · [arXiv](https://arxiv.org/abs/2606.14084)

25. #### [RoboGaze: Evaluating Robot World Models via Structured Vision-Language Analysis](https://arxiv.org/abs/2606.28385)

    M.L. Nguyen, N.T. Diep, H.K. Nguyen, M. Le, D.T. Le, H.H. Tran, D.D. Le, V.N. Duong, D. Sonntag, An Thai Le, D.M.H. Nguyen, V.A. Ngo, V.N. Tran

    RoboGaze uses structured vision-language analysis to classify errors in generated robot videos and identify when they occur. It requires no evaluator training.

    Under review · 2026 · [arXiv](https://arxiv.org/abs/2606.28385) · [Project page](https://robogaze-eval.github.io/) · [Code](https://github.com/cair-vinuni/RoboGaze)

26. #### [FINE: Future-Informed Navigation Encoding for Data-Efficient Vision-Language Navigation](https://arxiv.org/abs/2609.32855)

    K.H. Nguyen\*, H.P.Q. Nguyen\*, H.P. Nguyen, K.D. Binh, X.H. Nguyen, V.A. Ngo, D.M.H. Nguyen, H. Nguyen, An Thai Le

    FINE learns navigation from future observations in demonstrations. Tokens predict upcoming landmarks' appearance and geometry, and distinguish reached states from world-model alternatives. Gains on R2R-CE and RxR-CE grow with fewer demonstrations. This supervision adds no computation at deployment.

    Preprint · 2026 · [arXiv](https://arxiv.org/abs/2609.32855) · [Project page](https://finevln.github.io/)

27. #### [Retrieve to Act: Motion Primitive Graph Retrieval for Robust VLA Execution (R2A)](https://arxiv.org/abs/2610.37215)

    Pham Tri Quang, Bao-Ngoc Dao, Minh Duc Nguyen, Duc-Duy Nguyen, Hung T. Ho, Linh Dang Le, Trong-Bao Ho, Thien-Loc Ha, Tan Q. Nguyen, Tuan Dam, Vu N. Duong, Daniel Sonntag, Ngan Le, Khoa D. Doan, Jan Peters, An Thai Le, Duy M. H. Nguyen, Vien Anh Ngo

    R2A retrieves motions from successful demonstrations and blends them with a frozen VLA's actions. Tests use GR00T N1.7, π0.5, and StableVLA on LIBERO-Plus, RoboTwin2.0-Plus, and a real Aloha robot. Reported success gains reach 12 percentage points in simulation and 40 on hardware, without fine-tuning the base policies.

    Preprint · 2026 · [arXiv](https://arxiv.org/abs/2610.37215) · [Project page](https://retrieve2act.github.io/r2a/)

28. #### [AAC: Admissible-by-Architecture Differentiable Landmark Compression for ALT](https://arxiv.org/abs/2604.20744)

    An Thai Le, V.A. Ngo

    AAC compresses landmark heuristics for ALT graph search while preserving admissibility: the heuristic never overestimates the remaining path cost. We test it on road networks against fixed landmark-selection baselines.

    TMLR 2026 · [arXiv](https://arxiv.org/abs/2604.20744) · [Code](https://github.com/anindex/aac)

### 2025

1.  #### Representative paper: [Model Tensor Planning](https://openreview.net/forum?id=fk1ZZdXCE3) (MTP)

    An Thai Le, Khai Nguyen, Minh Nhat Vu, João Carvalho, Jan Peters

    MTP combines global candidates from layered graphs with local trajectory sampling for model predictive control. It batches simulator rollouts in MuJoCo XLA and is available in hydrax.

    TMLR 2025, ICLR 2026 (J2C) · ICRA 2025 RoboARCH Workshop (lightning talk) · [OpenReview](https://openreview.net/forum?id=fk1ZZdXCE3) · [Code](https://github.com/anindex/mtp) · [Workshop](https://sites.google.com/view/roboarch-icra25/schedule)

2.  #### Representative paper: [Global Tensor Motion Planning](https://sites.google.com/view/gtmp/home) (GTMP)

    An Thai Le, Kay Pompetzki, João Carvalho, Joe Watson, Julen Urain, Armin Biess, Georgia Chalvatzaki, Jan Peters

    GTMP searches a layered graph with tensor operations to produce batches of smooth spline paths in parallel. It provides probabilistic completeness under the paper's assumptions.

    IEEE RA-L 2025, ICRA 2026 · [DOI](https://doi.org/10.1109/LRA.2025.3575307) · [arXiv](https://arxiv.org/abs/2411.19393) · [Code](https://github.com/anindex/gtmp)

3.  #### Representative paper: [Motion Planning Diffusion: Learning and Adapting Robot Motion Planning with Diffusion Models](https://sites.google.com/view/motionplanningdiffusion) (MPD)

    João Carvalho, An Thai Le, Piotr Kicki, Dorothea Koert, Jan Peters

    MPD learns a diffusion prior over B-spline trajectories and guides sampling with task costs. Tests cover 2D tasks and 7-DoF arms, including real pick-and-place tasks learned from human demonstrations.

    IEEE T-RO 2025, AAAI 2026 (journal track) · [DOI](https://doi.org/10.1109/TRO.2025.3593109) · [arXiv](https://arxiv.org/abs/2412.19948) · [Code](https://github.com/joaoamcarvalho/mpd-splines-public)

4.  #### [FlowMP: Learning Motion Fields for Robot Planning with Conditional Flow Matching](https://arxiv.org/abs/2503.06135)

    Khang Nguyen, An Thai Le, T. Pham, M. Huber, Jan Peters, Minh Nhat Vu

    FlowMP uses second-order conditional flow matching to learn acceleration-aware motion fields for smooth robot trajectories.

    IROS 2025 · [arXiv](https://arxiv.org/abs/2503.06135) · [Code](https://github.com/mkhangg/flow_mp)

5.  #### [Machine Learning with Physics Knowledge for Prediction: A Survey](https://arxiv.org/abs/2408.09840)

    Joe Watson, Chen Song, Oliver Weeger, Theo Gruner, An Thai Le, Kay Hansel, Ahmed Hendawy, Oleg Arenz, Will Trojak, Miles Cranmer, Carlo D'Eramo, Fabian Bülow, Tanmay Goyal, Jan Peters, Martin W. Hoffmann

    We survey how physical knowledge informs model architectures, training objectives, and data. The review focuses on partial differential equations and open-source tools.

    TMLR 2025 · [arXiv](https://arxiv.org/abs/2408.09840)

### 2024

1.  #### [Structure-Aware E(3)-Invariant Molecular Conformer Aggregation Networks (ConAN)](https://arxiv.org/abs/2402.01975)

    D.M.H. Nguyen\*, N. Lukashina\*, T. Nguyen, An Thai Le, T. Nguyen, N. Ho, Jan Peters, D. Sonntag, V. Zaverkin, M. Niepert

    ConAN combines a 2D molecular graph with multiple 3D conformers through a differentiable fused Gromov–Wasserstein barycenter. Its molecular property predictions are invariant to rotations, translations, and reflections.

    ICML 2024 · [arXiv](https://arxiv.org/abs/2402.01975) · [Code](https://github.com/duyhominhnguyen/conan-fgw)

2.  #### [Dude: Dual Distribution-Aware Context Prompt Learning For Large Vision-Language Model](https://proceedings.mlr.press/v260/nguyen25c.html)

    D.M.H. Nguyen\*, An Thai Le\*, T.Q. Nguyen, N.T. Diep, T. Nguyen, D. Duong-Tran, Jan Peters, L. Shen, M. Niepert, D. Sonntag

    Dude adapts vision-language models from few examples. Unbalanced optimal transport aligns shared and LLM-generated class prompts with visual tokens.

    ACML 2024 · [PMLR](https://proceedings.mlr.press/v260/nguyen25c.html) · [arXiv](https://arxiv.org/abs/2407.04489)

3.  #### [Grasp Diffusion Network: Learning Grasp Generators from Partial Point Clouds with Diffusion Models in SO(3)×R³](https://sites.google.com/view/graspdiffusionnetwork) (GDN)

    João Carvalho, An Thai Le, Philipp Jahr, Qiao Sun, Julen Urain, Dorothea Koert, Jan Peters

    GDN generates grasp poses from partial point clouds using diffusion on SO(3)×R³ and collision-cost guidance. We test grasping from raw depth observations.

    Preprint · 2024 · [arXiv](https://arxiv.org/abs/2412.08398)

### 2023

1.  #### Representative paper: [Accelerating Motion Planning via Optimal Transport](https://sites.google.com/view/sinkhorn-step/) (MPOT)

    An Thai Le, Georgia Chalvatzaki, Armin Biess, Jan Peters

    MPOT optimizes batches of trajectories without gradients. The Sinkhorn Step coordinates waypoint updates through entropic optimal transport, with a Gaussian-process smoothness prior.

    NeurIPS 2023 · [arXiv](https://arxiv.org/abs/2309.15970) · [Code](https://github.com/anindex/mpot)

2.  #### [Hierarchical Policy Blending As Optimal Transport](https://sites.google.com/view/hipobot) (HiPBOT)

    An Thai Le, Kay Hansel, Jan Peters, Georgia Chalvatzaki

    HiPBOT uses unbalanced optimal transport to blend reactive Riemannian motion policies. A look-ahead layer selects the blending weights.

    L4DC 2023 · [Proceedings](https://proceedings.mlr.press/v211/le23a.html)

3.  #### [Motion Planning Diffusion: Learning and Planning of Robot Motions with Diffusion Models](https://sites.google.com/view/motionplanningdiffusion) (MPD)

    João Carvalho, An Thai Le, Mark Baierl, Dorothea Koert, Jan Peters

    This earlier version of MPD learns diffusion priors over robot trajectories. Denoising generates motion plans conditioned on a goal.

    IROS 2023 · [arXiv](https://arxiv.org/abs/2308.01557)

4.  #### [Learning to reason over scene graphs: a case study of finetuning GPT-2 into a robot language model for grounded task planning](https://arxiv.org/abs/2305.07716)

    Georgia Chalvatzaki, Ali Younes, Daljeet Nandha, An Thai Le, L.F.R. Ribeiro, I. Gurevych

    We fine-tune GPT-2 on scene graphs to turn long-horizon requests into subgoals for a planner. Tests use the ALFRED benchmark.

    Frontiers in Robotics and AI 2023 · [arXiv](https://arxiv.org/abs/2305.07716)

### 2022

1.  #### [Learning Implicit Priors for Motion Optimization](https://sites.google.com/view/implicit-priors/home) (StochGPMP)

    Julen Urain\*, An Thai Le\*, Alexander Lambert\*, Georgia Chalvatzaki, Byron Boots, Jan Peters

    We learn energy-based motion priors to initialize trajectory optimization or guide it through cost factors. Tests cover simulation and a real robot.

    IROS 2022 · [arXiv](https://arxiv.org/abs/2204.05369) · [Project page](https://sites.google.com/view/implicit-priors/home) · [Code](https://github.com/anindex/stoch_gpmp)

### 2021

1.  #### [Learning forceful manipulation skills from multi-modal human demonstrations](https://arxiv.org/abs/2109.04222)

    An Thai Le, Meng Guo, Niels van Duijkeren, L. Rozo, R. Krug, A.G. Kupcsik, M. Bürger

    The method learns insertion, sliding, and twisting from demonstrations of pose and force, using attractor-based impedance control. We demonstrate it in e-bike motor assembly.

    IROS 2021 · [arXiv](https://arxiv.org/abs/2109.04222)

2.  #### [Hierarchical Human-Motion Prediction and Logic-Geometric Programming for Minimal Interference Human-Robot Tasks](https://arxiv.org/abs/2104.08137)

    An Thai Le, P. Kratzer, S. Hagenmayer, M. Toussaint, Jim Mainprice

    We combine hierarchical human-motion prediction with logic-geometric replanning to reduce interference in shared workspaces. Tests use the MoGaze dataset.

    IEEE RO-MAN 2021 · [arXiv](https://arxiv.org/abs/2104.08137)

## Humanoid demos

At [VinRobotics](https://vinrobotics.net/), our researchers and engineers test learned locomotion, whole-body control, and physical interaction on real humanoids.

[Watch video: Compliant whole-body control](https://www.youtube.com/watch?v=TMuVG7YjUEk)

### Compliant whole-body control

CompliantWBC uses proprioceptive force estimates to coordinate the arms, torso, and legs of the ~70 kg VR-M3 during contact. The demos show guided interaction and cooperative tasks; the paper's passivity analysis is local and assumes fixed stiffness.

[Project page: Compliant whole-body control](https://compliantwbc.github.io/)

[Watch video: Perceptive stair locomotion](https://www.youtube.com/watch?v=rMq0qX4UDoY)

### Perceptive stair locomotion

VR-M3 (~70 kg) climbs real staircases of varying geometry at up to 0.6 m/s, including with a 5 kg payload. It uses a single onboard camera and a learned locomotion policy, with no LiDAR, motion capture, or teleoperation.

[Read article: Perceptive stair locomotion](https://vinrobotics.net/blog/perceptive-stair-locomotion)

[Watch video: Fast walking on VR-H3](https://www.youtube.com/watch?v=AIRifsVKtyc)

### Fast walking on VR-H3

VR-H3 (178 cm, 85 kg) uses an RL locomotion policy trained with phase-aware gait rewards and domain randomization. The training curriculum increases difficulty only after the policy learns to balance at each level.

[Read article: Fast walking on VR-H3](https://vinrobotics.net/blog/vin-robotics-humanoid-robots-reaching-human-level-walking-speed)

[Watch video: Built from the motor up](https://www.youtube.com/watch?v=dlNunvCxXcw)

### Built from the motor up

In-house actuators and real-time EtherCAT communication support learned locomotion on our humanoid platform. The team reports walking at 1.5 m/s, reaching 1.8 m/s with controllers trained on human motion data.

[Read article: Built from the motor up](https://vinrobotics.net/blog/humanoid-heartbeat)

[Watch video: Real-world football](https://www.youtube.com/watch?v=k_bzCFVfQ_k)

### Real-world football

An outdoor football game with people: passing, running, and celebrating goals together.

[Read article: Real-world football](https://vinrobotics.net/blog/a-new-milestone-in-proud-times)

[Watch video: Global debut](https://www.youtube.com/watch?v=rSVRZtWi1Yc)

### Global debut

Platform preview for Computex and ICRA 2026: whole-body teleoperation, dynamic payload handling, MPC + RL locomotion, and perception-action learning.

[Read article: Global debut](https://vinrobotics.net/blog/vin-robotics-humanoid-platform-global-debut)

## Join us

Our group is based at VinUniversity in Hanoi, with close ties to VinRobotics and IAS at TU Darmstadt. If you enjoy understanding algorithms and testing them on robots, I'd like to hear from you.

### How we work together

I help students develop their own research judgment. We work through the mathematics and code together, then design experiments that test our assumptions. There is room to make mistakes, rethink an idea, and pursue your own direction. I encourage students to question mine, too.

### Open positions and projects

- PhD and MSc research at [VinUniversity](https://vinuni.edu.vn/).
- Research residencies and internships at [VinRobotics](https://vinrobotics.net/) (Foundation AI).
- MSc theses and co-advised PhD projects at [IAS, TU Darmstadt](https://www.ias.informatik.tu-darmstadt.de/).

### How to apply

- Email <an@robot-learning.de> with the subject `[Join][<track>] <Your Name>`. Use PhD, MSc, residency, or internship as the track.
- Attach a CV and a transcript, and link to code you wrote.
- Add a short paragraph about one of our papers (under 300 words). Describe an idea, a question, or an experiment you would like to try.

I read emails with this subject line first, but may not be able to reply to every message. Thanks for your patience.

## Group

The students, residents, alumni, and collaborators behind our research. Read about [how we work together and how to join](https://anindex.github.io/#join).

### MSc students

Dinh Van The Long [Learning under partial observability](https://proceedings.mlr.press/v270/nguyen25b.html)

### VinRobotics residents

Trinh Thi Cuc [Whole-body compliance](https://anindex.github.io/#pub-compliantwbc)

Dang Truong Duy [Payload-robust locomotion](https://anindex.github.io/#pub-tactful)

Le Anh Chien [Humanoid locomotion & control](https://anindex.github.io/#pub-tactful)

Nguyen Viet Duong Neuromorphic learning

Ly Phuc Thanh [Perceptive humanoid locomotion](https://anindex.github.io/#pub-tactful)

Do Tan Dung [Humanoid behavior models](https://anindex.github.io/#pub-crossbfm)

Ho Thinh Hung [VLA quantization & deployment](https://anindex.github.io/#pub-foldquantvla)

Ha Thien Loc [Equivariant robot policies](https://anindex.github.io/#pub-equivla)

Nguyen Quang Tan [Equivariance & action generation](https://anindex.github.io/#pub-equivla)

Ho Trong Bao [Asynchronous robot policies](https://anindex.github.io/#pub-paint)

Nguyen Dang Khanh [Efficient VLA inference](https://anindex.github.io/#pub-vlasimd)

Le Van Minh [World models & synthetic data evaluation](https://anindex.github.io/#pub-robogaze)

Nguyen Minh Hoang [Reinforcement learning & agents](https://mhngu23.github.io/mhnguyen-ai.github.io/)

Dao Duc Thinh [Egocentric VLA learning](https://thinh-dao.github.io/)

Truong Minh Hoang [CPU inference & SIMD kernels](https://anindex.github.io/#pub-vlasimd)

Nguyen Truong Chinh [VLA inference runtimes](https://anindex.github.io/#pub-vlacpp)

### Alumni

Phuong Tuan Dat [Cross-embodiment robot learning](https://anindex.github.io/#pub-crossbfm)

Magnus Dierking [Contact-rich model predictive control](https://magnusdierking.github.io/)

Caio Freitas [Graph diffusion for imitation learning](https://caio-freitas.github.io/)

Qiao Sun [Geometric diffusion for grasping](https://anindex.github.io/#pub-graspdiffusion)

Denis Andrić [Learning Hamiltonian dynamics](https://www.ias.informatik.tu-darmstadt.de/Off-Campus/AnThaiLe)

Sebastian Zach [Reactive motion generation](https://www.ias.informatik.tu-darmstadt.de/Off-Campus/AnThaiLe)

### Collaborators

[Zachary Kingston](https://commalab.org) Purdue University

[Meng Guo](https://mengguo.github.io/personal_site/) Peking University

[Nguyen Ho Minh Duy](https://duyhominhnguyen.github.io) DFKI

[Ngo Anh Vien](https://vienngo.github.io) VinRobotics

[Jan Peters](https://www.ias.informatik.tu-darmstadt.de/Member/JanPeters) TU Darmstadt

[Georgia Chalvatzaki](https://pearl-lab.com/people/georgia-chalvatzaki/) TU Darmstadt

[João Carvalho](https://joaoamcarvalho.github.io/) DFKI

[Viet T. Nguyen](https://vietngth.github.io/) University of Würzburg

## Open source

Research code and deployment tools developed with students and collaborators, including our releases at [CAIR, VinUniversity](https://github.com/cair-vinuni). Start with [GTMP](https://github.com/anindex/gtmp) or [MPOT](https://github.com/anindex/mpot) for planning, [vla.simd](https://github.com/cair-vinuni/vla.simd) for CPU policy inference, or [warpax](https://github.com/anindex/warpax) for numerical relativity. Questions and bug reports are welcome; reproducible examples help us investigate.

[cair-vinuni/vla.simd](https://github.com/cair-vinuni/vla.simd): C++ inference for VLA policies with native SIMD kernels on x86, Apple Silicon, and Raspberry Pi 5. (C++)  
[cair-vinuni/FoldQuantVLA](https://github.com/cair-vinuni/FoldQuantVLA): Native INT4 and INT8 inference for VLA models through consistent offline folding, with deployment on Jetson AGX Orin. (Python)  
[cair-vinuni/FOCA](https://github.com/cair-vinuni/FOCA): Adapts VLA policies from few demonstrations by learning from future interactions and goals. ICML 2026. (Python)  
[cair-vinuni/CLP_VLA](https://github.com/cair-vinuni/CLP_VLA): Prunes redundant VLA layers before fine-tuning, with training and evaluation scripts for robot manipulation. (Jupyter Notebook)  
[cair-vinuni/RoboGaze](https://github.com/cair-vinuni/RoboGaze): Diagnoses failures in videos generated by robot world models using structured vision-language analysis. (Python)  
[cair-vinuni/paint-action-chunking](https://github.com/cair-vinuni/paint-action-chunking): Project materials for PAINT: connecting action chunks through initial noise selection without retraining. CoRL 2026.  
[VinRobotics/vla.cpp](https://github.com/VinRobotics/vla.cpp): C++ runtime for onboard inference with eleven vision-language-action models. (C++)  
[anindex/mpot](https://github.com/anindex/mpot): Gradient-free trajectory optimization with optimal transport in PyTorch. NeurIPS 2023. (Python)  
[anindex/mtp](https://github.com/anindex/mtp): Global and local trajectory sampling for model predictive control in JAX. TMLR 2025 & ICLR 2026. (Python)  
[anindex/gtmp](https://github.com/anindex/gtmp): Parallel motion planning with layered graphs and tensor operations in JAX. RA-L 2025 & ICRA 2026. (Python)  
[anindex/polystep](https://github.com/anindex/polystep): Optimal-transport training for non-differentiable networks using only forward evaluations. TMLR 2026. (Python)  
[anindex/warpax](https://github.com/anindex/warpax): Vectorized auto-diff in JAX for checking energy conditions in warp drive spacetimes. CQG 2026. (Python)  
[anindex/penrose_process](https://github.com/anindex/penrose_process): Rocket-flyby simulations testing energy extraction and escape around rotating black holes. Phys. Rev. D 2026. (Python)  
[anindex/aac](https://github.com/anindex/aac): Learned compression of landmark heuristics for A\* search, preserving admissibility by construction. (Python)  
[VinRobotics/vinrobotics_mjlab](https://github.com/VinRobotics/vinrobotics_mjlab): RL training pipeline for high-payload humanoid locomotion, built on MuJoCo Warp. (Python)  
[VinRobotics/model-quantization-recipes](https://github.com/VinRobotics/model-quantization-recipes): Quantization recipes for language and speech models, with deployment checks. (Python)  
[anindex/note_model_opt](https://github.com/anindex/note_model_opt): Notes and runnable examples on model deployment.  
[vincekurtz/hydrax (contributor)](https://github.com/vincekurtz/hydrax): Sampling-based model predictive control on the GPU with JAX/MJX. Includes MTP. (Python)  

## Teaching & mentoring

I start with a small example, work through the mathematics, then implement it to test the assumptions. Our [student projects and mentoring](https://anindex.github.io/#join) follow the same approach: learn to ask precise questions and judge the evidence for yourself.

### Courses at TU Darmstadt

- [Statistical Machine Learning](https://www.ias.informatik.tu-darmstadt.de/Teaching/StatisticalMachineLearning) · Teaching assistant Summer 2023, Winter 2023/24,  
  Summer 2024, Winter 2024/25
- [Probabilistic Methods in Computer Science](https://www.ias.informatik.tu-darmstadt.de/Teaching/PROMI) Winter 2024/25
- Robot Learning Integrated Project / Expert Lab / Mechatronics Winter 2024/25
- [Reinforcement Learning](https://www.ias.informatik.tu-darmstadt.de/Teaching/ReinforcementLearning) · Teaching assistant Summer 2022

### Open learning resources

- [Designing Planning Algorithms for the Era of Parallelism](https://anindex.github.io/slides/roboarch/) · talk slides on tensor search and parallel motion planning.
- [Model deployment notes](https://github.com/anindex/note_model_opt) · examples to run and modify.

## Community & service

### Editorial and reviewing roles

Associate Editor  
[IEEE RA-L](https://www.ieee-ras.org/publications/ra-l), [ICRA 2027](https://www.ieee-ras.org/conferences-workshops/fully-sponsored/icra)

Area Chair  
[ICML 2027](https://icml.cc/), [ICLR 2027](https://iclr.cc/), [CoRL 2026](https://www.corl.org/), RLC 2024

Conference reviewer  
IROS, ICRA, RSS, L4DC, NeurIPS (Top Reviewer 2026), ICML, ICLR, AAAI

Journal reviewer  
IEEE RA-L, IEEE T-RO, Robotics and Autonomous Systems, Neurocomputing, TMLR, Frontiers in Robotics and AI

### Selected invited talks

- **Designing Planning Algorithms for the Era of Parallelism** · [ICRA Workshop on Robot Architecture](https://sites.google.com/view/roboarch-icra26/home) and the [2nd Next-Gen Robot Learning Symposium](https://sites.google.com/view/robot-learning-symposium), 2026. [Slides](https://anindex.github.io/slides/roboarch/).
- **Tensor search methods for motion planning** · RMIT University, Rice University, and HUST, 2025.

## Experience & education

1.  Oct 2025–present  
    Hanoi, Vietnam

    ### Assistant Professor

    [VinUniversity](https://vinuni.edu.vn/)

    Building a research group on efficient learning and planning for robot locomotion and manipulation.

2.  Oct 2025–present  
    Darmstadt, Germany

    ### Visiting Professor

    [TU Darmstadt](https://www.tu-darmstadt.de/)

    Co-advising MSc and PhD students in robot learning at IAS.

3.  Aug 2025–present  
    Hanoi, Vietnam

    ### Director of Foundation AI

    [VinRobotics](https://vinrobotics.net/)

    - RL stack for high-payload humanoid locomotion
    - Humanoid VLA architecture and training recipe
    - Model optimization and edge-deployment toolchain

4.  2022–2025  
    Darmstadt, Germany

    ### PhD in Computer Science

    [TU Darmstadt, Intelligent Autonomous Systems (IAS)](https://www.ias.informatik.tu-darmstadt.de/)

    Thesis: *Tensor Search Methods for Vectorizing Motion Planning*, supervised by [Prof. Jan Peters](https://www.ias.informatik.tu-darmstadt.de/Team/JanPeters).

5.  2019–2021  
    Stuttgart, Germany

    ### MSc in Information Technology

    [Universität Stuttgart](https://www.uni-stuttgart.de/)

    Thesis: *Learning task-parameterized Riemannian motion policies*, supervised by [Dr. Jim Mainprice](https://sites.google.com/site/jimmainprice/home) and [Dr. Meng Guo](https://mengguo.github.io/personal_site/). Graduated first class. Info-Preis for the best diploma, Sony Research Award, Deutschlandstipendium.

6.  May–Dec 2020  
    Renningen, Germany

    ### Research Intern

    [Bosch Center for Artificial Intelligence](https://www.bosch.com/research/bcai/)

    Forceful imitation learning for e-bike assembly, hosted by [Dr. Meng Guo](https://mengguo.github.io/personal_site/) in the robotics team.

7.  Nov 2019–Apr 2020  
    Stuttgart, Germany

    ### Research Assistant

    [HLRS (High-Performance Computing Center Stuttgart)](https://www.hlrs.de/)

    Configured and maintained HPC systems.

8.  2015–2019  
    Frankfurt, Germany

    ### BEng in Electrical Engineering and Information Technology

    [Frankfurt University of Applied Sciences](https://www.frankfurt-university.de/)

    Thesis: *Approaches to solve the kidnapped robot problem*. Graduated first class. DAAD, AmCham and eSilicon scholarships.

9.  Jan–May 2017  
    Ho Chi Minh City, Vietnam

    ### Engineering Intern

    [Intel Corporation](https://www.intel.com/)

    Designed systems to analyze high-volume manufacturing unit-test data and supported quality validation and reporting on the Thunderbolt line.

## Contact

<an@robot-learning.de>

VinUniversity, Hanoi, Vietnam

## Elsewhere

- [VinUniversity profile](https://vinuni.edu.vn/people/le-thai-an/)
- [CAIR, VinUniversity](https://cair.vinuni.edu.vn/people/)
- [IAS, TU Darmstadt](https://www.ias.informatik.tu-darmstadt.de/Team/AnThaiLe)
- [Experience & education](https://anindex.github.io/#experience)
- [Humanoid demos](https://anindex.github.io/#demos) · [Teaching](https://anindex.github.io/#teaching) · [Community](https://anindex.github.io/#service)

## This site

Last updated November 2026.

© 2026 An Thai Le