Publications

TL;DR: VICAR learns interactive humanoid skills from third-person human videos while preserving world-frame contact poses, motion speed, and collision clearance, so contact-rich skills survive embodiment mismatch and runtime scene changes.

TL;DR: RPG reconstructs practice tasks from demonstrations, uses guided practice to improve shared skills, and then goes real. Across 22 manipulation tasks, RPG improves from 28.6% to 95.0% over 15 practice rounds and achieves 30/30 successful trials across drawer closing, towel folding, and bimanual handover.

TL;DR: One human demo. Any multi-fingered hand. Zero-shot sim-to-real visuomotor policy. Morphometric Imitation turns reconstructed hand-object interactions into policies that reach 89.3% zero-shot success across 300 real-world trials on 30 objects.

TL;DR: Success-Guided Sampling (SGS) trains on the task configurations that are neither too easy nor too hard for the current policy. With it, reinforcement learning scales much more effectively, to over a million simulated robots at once, in legged locomotion and contact-rich manipulation.

TL;DR: ENPIRE enables coding agents to conduct physical autoresearch to autonomously hill-climb on real-world robot tasks. The agent designs the environment and reward, then runs experiments in parallel.

TL;DR: CaP-X evaluates frontier LLM performance on robotics tasks by providing agentic harnesses. Agents write code to interact with the environment, get feedback, and improve via RL and skill libraries. These models show far superior generalization compared to frontier robotics models like VLAs.

TL;DR: DoorMan proposes a teacher-student-bootstrap framework for challenging humanoid loco-manipulation tasks such as door opening. Trained as an RGB policy purely in simulation, it is up to 31.7% times faster than human in the real world.

TL;DR: VIRAL investigates the scaling law of visual sim-to-real. We find the right recipe to enjoy the free lunch of simulation: zero-shot, robust, continuous real-world deployment.

TL;DR: SONIC is a general humanoid whole-body motion tracker supporting various control modes.

TL;DR: (Probe, Learn, Distill) is a plug-and-play recipe for Vision-Language-Action (VLA) post-training. It is model agnostic, supporting both autoregressive and diffusion architectures, and can push success rates to 99%.

TL;DR: DreamControl leverages a generative model trained on offline human motion data to synthesize diverse human motion trajectories for a wide range of tasks. These trajectories are re-targeted to the humanoid, and used to train task-specific RL policies that can be deployed both in simulation and on some real-world tasks.

TL;DR: Model based planner + reference motion tracking enables robust real-world Ping Pong play (continuiously up to 106 hits!)

TL;DR: LEGO improves generalized grasping capability of off-the-shelf VLAs like π0-Fast by learning from grasping toys randomly generated from simple primitives. We use a novel object-centric representation to ensure the visual features stay grounded in the transfer.

TL;DR: LeVERB is the first latent whole-body humanoid VLA. We introduce a latent vocabulary as an interface between vision-language and whole-body action to enable expressive task specification and interpolatable execution.

TL;DR: ARM4R is an Autoregressive Robotic Model that leverages low-level 4D Representations learned from human video data to yield a robotics model that has stronger spatial and temporal understandings.

TL;DR: LATTE-MV is a scalable system that processes over 800 hours of YouTube table tennis footage to reconstruct 27 hours of clean gameplay. This dataset is then used to train an anticipatory controller that predicts the player's next shot, enabling the robot to pre-position itself and perform better against high-speed rallies.

TL;DR: DIAL-MPC is the first training-free method achieving real-time whole-body torque control using full-order dynamics.