Action Map Policy casts closed-loop 3D manipulation as dense pixel classification, preserving multimodal action distributions while predicting a full action chunk in one forward pass.

Pix2Act represents manipulation as continuous image-space keypoint trajectories, enabling equivariant augmentation and robust reconstruction of closed-loop 3D actions.

Retriever is a programming model and runtime that we use to build hierarchical pipelines for long-horizon, partially observable robot manipulation.

A benchmark for evaluating observation space shift challenges in long-horizon robotic tasks.

We use vision-language foundation models as uncertainty estimators inside a symbolic belief-space planner, enabling robots to plan to perceive and act under partial observability.

This thesis studies how structured abstractions, including symmetry, compositionality, and belief-state representations, can make learning and planning more efficient for embodied decision-making.

GEM combines patch-level semantic maps, language-conditioned visual relevancy, and equivariant policy learning for sample-efficient open-vocabulary pick-and-place manipulation.

We enable a robot to rapidly and autonomously specialize parameterized skills by planning to practice them. The robot decides what skills to practice and how to practice them. The robot is left alone for hours, repeatedly practicing and improving.

We have developed ThinkGrasp, a plug-and-play vision-language grasping system for heavy clutter environment grasping strategies.

We study the problem of action sampling and propose a method to incorporate equivariance properties to the action sampling procedure.

This work learns to navigate end-to-end with map input, aiming to generalize to novel map layouts in zero-shot.

We have developed ThinkGrasp, a plug-and-play vision-language grasping system for heavy clutter environment grasping strategies.

We study E(2) Euclidean equivariance in navigation on geometric graphs and develop message passing network to solve it.

We formulate how differentiable planning algorithms can exploit inherent symmetry in path planning problems, named SymPlan, and propose practical algorithms.

We differentiate through the Bellman fixed point to decouple forward planning from backward computation, enabling a flexible forward budget and constant backward cost in planning horizon.

We study sample-efficient modeling of invariant drag coefficients with equivariant networks.

We train an equivariant network for pose prediction from single 2D image by using induced and restricted representations.

We study whether Euclidean symmetry can help in reinforcement learning and planning, which models the geometric transformations between reference frames of robots.

We formulate compositional generalization in object-oriented world modeling, and propose a soft and efficient mechanism for practice.

We learn latent representations enforced with known transformation laws (group action), and apply this idea on equivariant latent world modeling.

We propose a parameterized action reinforcement learning algorithm to improve the performance of match plan generation in Bing search.

We present a deep imitation learning framework for robotic bimanual manipulation in a continuous state-action space.

This work aims to improve the zero-shot generalization performance in map-based navigation on novel layouts.

This work proposes a reinforcement learning framework for compositional object-oriented environments.