Richard Li

I am a PhD student at MIT CSAIL, where I work on robotics and machine learning.

Projects

My research interest lies in finding a scalable path to robotics foundation models. I believe this will require solving cross-embodiment learning—especially transferring knowledge from human videos —as well as developing reinforcement learning methods that scale to large, real-world robotics datasets.

	Learning More from Less: Reinforcement Learning from Hindsight Iris Xu, Sunshine Jiang, John Marangola, Nitish Dashora, Richard Li, Thomas Liu, Zexue He, Yuheng Zhi, Alex Pentland, Pulkit Agrawal, Zhang-Wei Hong arXiv preprint, 2026 arXiv Relabeling failed VLA rollouts with hindsight instructions and rewards from a VLM yields 5× better sample efficiency than standard RL for vision-language-action policies.
	What Matters When Cotraining Robot Manipulation Policies on Everyday Human Videos? Richard Li, Aditya Prakash, Andrew Wen, Saurabh Gupta, Yilun Du, Pulkit Agrawal RSS 2026: Data-Centric Robotics Workshop project page / arXiv Hand pose quality and embodiment-specific network specialization are key to enabling transfer from everyday videos to robots.
	Prompt-Driven Exploration for VLA Policies Sunshine Jiang, John Marangola, David Zhang, Raghuram Kowdeed, Ruiyang Luo, Nitish Dashora, Richard Li, Pulkit Agrawal, Zhang-Wei Hong ICRA 2026: VLA Pipelines for Real Robots Workshop project page / paper Exploring in language space: a VLM iteratively refines task prompts from rollout observations, enabling VLA policies to bootstrap reinforcement learning from zero-reward starts.
	Bimanual 3D Hand Motion and Articulation Forecasting in Everyday Images Aditya Prakash, Richard Li, David Forsyth, Saurabh Gupta 2025 project page / code Forecasting bimanual 3D hand motion and articulation from a single image using a diffusion-based lifting and forecasting pipeline.
	Stable Object Reorientation using Contact Plane Registration Richard Li, Carlos Esteves, Ameesh Makadia, Pulkit Agrawal ICRA 2022 project page / video / code Predicting contact points with a CVAE and plane segmentation improves object generalization and handles multimodality.
	Contact-Aware Lyapunov Controller Design via Alternating Optimization Richard Li, Timur Garipov 2022 paper / video Synthesizing Lyapunov controllers through contact with alternating optimization.
	Towards Practical Multi-object Manipulation using Relational Reinforcement Learning Richard Li, Allan Jabri, Trevor Darrell, Pulkit Agrawal ICRA 2020, ICML 2020: Bridge Between Perception and Reasoning Workshop project page / video / code Multi-object, long-horizon manipulation can be autonomously learned using a curriculum and graph neural network architecture.

Template.