Email: juntao.ren@stanford.edu
Juntao Ren
I'm a Ph.D. student at Stanford University advised by Shuran Song and part of the Robotics and Embodied AI Lab (REAL). I enjoy thinking about how we can enable robots to learn and reason over time in the physical world. I'm grateful to be supported by the Knight-Hennessy Fellowship and the NSF Graduate Research Fellowship.
I completed my undergrad in Computer Science and Mathematics at Cornell University, where I'm grateful to have worked with Sanjiban Choudhury. I've also spent time at 1X Technologies working on World Models.
Please feel free to reach out!
Publications
* indicates equal contribution
NEEDLEWORK: Offline Rewriting of Robot Data with Verified Local Stitches
TL;DR — We study how to get stitching to work on robot data, and more importantly, why these connections improve behavior beyond demonstration data.
A Smooth Sea Never Made a Skilled SAILOR: Robust Imitation via Learning to Search
TL;DR — We train both world and reward models from demonstration data, giving the agent the ability to reason about how to recover from mistakes at test time.
Motion Tracks: A Unified Representation for Human-Robot Transfer
TL;DR — We propose a unified action space by representing actions as 2D trajectories on an image, enabling robots to directly imitate from cross-embodiment datasets.
Hybrid Inverse Reinforcement Learning
Paper · Website · Code · Video
TL;DR — We show that training on both expert and learning data can provably speed up interactive imitation learning in the absence of rewards, for both model-free and model-based algorithms.
MOSAIC: A Modular System for Assistive and Interactive Cooking.
TL;DR — We build a top-down modular system to allow for human-robot collaboration within the kitchen using high-level planners, low-level visuomotor policies, and human-motion forecasting.
Test-Time Distribution Normalization for Contrastively Learned Visual-language Models.
TL;DR — We prove that subtracting the mean of the image and language embeddings at test-time from each sample better aligns with the training-objective and improves performance.
Projects
1X World Model: Evaluating Bits, not Atoms
