We introduce Distribution Contractive Reinforcement Learning (DICE-RL), a framework that uses reinforcement learning (RL) as a "distribution contraction" operator to refine pretrained generative robot policies. DICE-RL turns a pretrained behavior prior into a high-performing "pro" policy by amplifying high-success behaviors from online feedback. It enables mastery of complex long-horizon manipulation skills directly from high-dimensional pixel inputs, both in simulation and on a real robot.

We introduce Latent Policy Barrier, a framework for robust visuomotor policy learning. LPB treats the latent embeddings of expert demonstrations as an implicit barrier separating safe, in-distribution states from unsafe, out-of-distribution (OOD) ones. Our approach decouples the role of precise expert imitation and OOD recovery into a base diffusion policy and a dynamics model. At inference time, the dynamics model predicts future latent states and optimizes them to stay within the expert distribution.

We introduce a method that automatically generates reward functions for agents to learn new tasks using only a text description of the task goal and the agent’s visual observations, by leveraging feedback from vision language foundation models (VLMs).

We introduce a method that leverages both vision and force modalities for robot-assisted dressing. Our method combines the vision-based policy, trained in simulation, with the force dynamics model, learned in the real world to achieve better dressing performance and safety for the user.

We develop, for the first time, a robot-assisted dressing system that is able to dress different garments on people with diverse body shapes and poses from partial point cloud observations, based on a single reinforcement learning policy.

We propose a dedicated algorithm and accelerator co-design framework dubbed ViTCoD to accelerate ViTs. On the algorithm level, we prune and polarize the attention maps to have either denser or sparser patterns. On the hardware level, we develop an accelerator to coordinate the denser/sparser workloads for higher hardware utilization.

We discover for the first time that both efficient DNNs and their lottery subnetworks can be identified from a supernet and propose a two-in-one training scheme with jointly architecture searching and parameter pruning to identify them.

We propose a method that leverages human guidance for high DOF robot motion planning in partial observable environments. We project the robot’s continuous configuration space to a discrete task model and utilize inverse RL to learn motion-level guidance from human critiques.