RSS SemRob 2026

About
How can robots learn not just to act, but to understand what their actions mean?

RSS SemRob 2026 explores semantic reasoning and goal understanding as foundations for more robust, adaptive, and generalizable robot learning. The workshop brings together researchers across robot learning, embodied AI, planning, cognitive robotics, neuro-symbolic AI, vision, language, and multimodal learning to discuss how robots can learn task-relevant representations, reason over objects and relations, and ground semantic knowledge in real-world behavior.

Topics include semantic and multisensory representations for policy learning, hybrid neuro- symbolic architectures, structured world models, robot foundation models, memory and metacognitive reasoning, safety-aware learning and control, efficient reasoning on resource-constrained robotic platforms, as well as cross-embodiment transfer and human-motion retargeting.

SemRob will be held at RSS 2026 in Australia and will feature invited talks, contributed presentations, posters, and a structured debate on how semantic structure can help robots learn more efficiently, generalize across tasks and environments, and act reliably in uncertain, dynamic worlds.

Schedule & Speakers (Morning, Half-day workshop) July 17th, 2026 - Sydney, Australia

Time		Info
07:50		Organizers: Introductory Remarks
08:00		Keynote: William Zhi University of Sydney Title: TBD Abstract coming soon.
08:30		Keynote: Hang Zhao Tsinghua University + Galaxea AI Title: Autoregressive VLA: Reasoning and Acting in One Autoregressive Stream Can a robot reason about what to do and execute how to do it within a single autoregressive process? In this talk, I will present our work on autoregressive vision-language-action models, where high-level reasoning and low-level action generation are modeled in one unified token stream. This formulation allows semantic understanding, task decomposition, and continuous control to share the same modeling interface and training objective, rather than being handled by separately designed action experts. I will discuss the evolution of our action tokenizers and models, including FASTer, ActionCodec and Galaxea G0.5.
09:00		Keynote: Jean Oh Carnegie Mellon University + Lavoro AI Title: Safe Physical AI under the Curse of Rarity: From Self-Driving to Aviation Abstract coming soon.
09:30		Keynote: Jesse Zhang University of Washington Title: Learn Less, Borrow More: Generalization doesn’t have to be re-learned — it can be borrowed Abstract coming soon.
10:00		Keynote: Andreea Bobu Massachusetts Institute of Technology Title: Reading Between the Lines: Using Language Models to Amplify Human Data in Robot Learning Human-in-the-loop robot learning faces a fundamental data challenge that general machine learning doesn't: unlike settings where we can collect massive offline datasets, robots must learn from limited, real-time human interactions. This creates a critical bottleneck: we need methods that can make the most of limited human input, or, in other words, that can learn a lot from a little. The challenge is that humans are imperfect communicators of their own intent: language instructions are often ambiguous, demonstrations can be incomplete or overfit to a specific setting, and physical corrections have multiple valid interpretations. Our key insight is that human feedback is shaped by context in predictable ways, and that modeling that context turns ambiguous, incomplete feedback into a rich signal about underlying intent. In this talk, I will discuss three sources of such structure: 1. human feedback modalities are incomplete in complementary ways, and that complementarity is itself a source of signal; 2. human feedback is only interpretable relative to how humans represent the world, and learning that representation is as important as learning the reward; and 3. behind any specific human input lies a higher-level intent that generalizes far beyond the situation in which it was expressed. Together, these directions show that understanding the structure of human communication — rather than simply collecting more of it — is the key to efficient, generalizable, human-aligned robot learning.
10:30		Spotlight talks
11:00		Coffee Break + Poster session
11:30		Panel discussion Panelists: William Zhi, Hang Zhao, Jean Oh, Jesse Zhang, Andreea Bobu
12:25		Organizers: Closing Remarks, Best paper

Targeted Topics

We especially welcome work that investigates how semantic structure can improve the data efficiency, generalization, adaptability, safety, and interpretability of learned robot policies:

Semantic representations for generalizable robot learning, including object-centric representations, affordances, skill abstractions, geometric task-space primitives, pre-/post-conditions, constraints, and action tokenization
Vision-language-action models, robot foundation models, and generalist policies that connect semantic understanding with embodied control
Hybrid neural-symbolic architectures that combine learned representations with structured scene representations, knowledge graphs, scene graphs, task graphs, or procedural reasoning
World models and dynamics models for robot learning, including predictive simulation, synthetic data generation, policy training, planning, evaluation, and sim-to-real transfer
Cross-embodiment robot learning, including shared action spaces, embodiment-invariant task representations, morphology-aware policies, and transfer across robot platforms
Test-time adaptation, policy steering, online policy guidance, and failure recovery for robots operating in uncertain or changing environments
Human-in-the-loop robot learning, including online policy supervision, preference feedback, corrective demonstrations, interactive teaching, and behavior repair
Semantic memory mechanisms for storing, retrieving, and reusing experience across tasks, environments, and interaction histories
Metacognitive and self-reflective mechanisms for robot learning, including uncertainty estimation, calibration, self-critique, chain-of-thought-style reasoning, agentic reasoning structures, and confidence-aware control
Safety-aware semantic reasoning for learned robot policies, including constraint satisfaction, risk-aware decision-making, verification, monitoring, and intervention mechanisms
Data-efficient and scalable robot learning methods that exploit semantic structure, task decomposition, compositionality, or foundation-model priors
Benchmarks, datasets, and evaluation protocols for semantic generalization, long-horizon behavior, cross-task transfer, and real-world robustness

Submission Guidelines

RSS SemRob 2026 suggests 4+N or 8+N paper length formats — i.e., 4 or 8 pages of main content with unlimited additional pages for references, appendices, etc.

Submission Site https://openreview.net/group?id=roboticsfoundation.org/RSS/2026/Workshop/SemRob

We will accept the official LaTeX or Word paper templates, provided by RSS 2026. Authors should upload their submission as a PDF file to OpenReview.

Our review process will be double-blind, following the RSS paper submission policy for Science/Systems papers.

All accepted papers will be invited for poster presentations; the highest-rated papers, according to the Technical Program Committee, will be given spotlight presentations. We will highlight the Best Paper Award during the closing remarks at the workshop event.

Important Dates

Submission deadline (extended): ~~8 June 2026~~
18 June 2026
, 23:59 AOE.
Author Notifications: 29 June 2026, 23:59 AOE.
Camera Ready: 5 July 2026, 23:59 AOE (firm).
Workshop: 17 July 2026.

Dual Submission Policy

We welcome new, under-review, or recently published work. Accepted papers will be posted on this website but are non-archival and won't be included in any official proceedings.