TL;DR. (Dexterous Manipulation Learning from Human Video) Built a scalable data engine to retrieve human demonstrations from internet videos through 3D hand-motion matching. Improved manipulation success in Isaac Gym simulation and zero-shot sim-to-real transfer to the LEAP Hand.

TL;DR. (Understanding Representations in Robot Foundation Models) Offered a representation-learning view of embodied model success. Found that models with higher task success show stronger alignment between their vision and language representations and motion-trajectory representations. Developed an offline proxy based on this finding to assess VLAs and WAMs before closed-loop evaluation, diagnose models, and explain empirically observed behaviors.

TL;DR. (RL for Multimodal Reasoning) Showed that RL post-training on simple arcade games improves multimodal LLMs' out-of-domain math and spatial reasoning without worked solutions or diagrams during that post-training stage.

TL;DR. (Agent Memory and Self-Evolution) Built a self-play framework that distills agent trajectories into reusable insights stored in persistent memory and evolves agent prompts, improving performance and reducing run-to-run variance across five text games.

TL;DR. (Efficient LLM Reinforcement Learning) Showed that GRPO tolerates much staler rollouts than previously assumed when update instability is controlled. Reused rollouts in a few large generation and optimization stages, matching standard GRPO's reasoning performance at lower training cost.

TL;DR. (Multimodal Data Curation) Built an automated grounding, annotation, and filtering multimodal data pipeline for medical pretraining, covering 25 million images across 10 modalities.

TL;DR. (Image Segmentation) Used local and global aggregation in a hierarchical vision transformer to improve part and object segmentation.

TL;DR. (Long-Context Image Generation) Improved consistency in long-story generation with a training-free method that refines story images using global reference context from the previous iteration.