Robotics • Machine Learning • Generalist Robots
I am a PhD student at Princeton University, advised by Prof. Anirudha Majumdar. I also work closely with Prof. Dhruv Shah. I am currently a research intern at Physical Intelligence. I am interested in building generalist robots.
Email: lihanzha [at] princeton [dot] edu
TL;DR: We introduce EgoLAP, a vision-language-action pre-training framework that learns from egocentric human and robot trajectories through shared language actions and motion-level reasoning. EgoLAP transfers human experience to robot control, achieving 80.1% mean real-world task progress on an unseen robot configuration and a 2.3× gain over alternative action representations.
TL;DR: We introduce EmbodiedSWE, a framework for using coding agents to solve long-horizon, dexterous robotics tasks and turn verified solutions into scalable training data. EmbodiedSWE-Bench evaluates agents across 28 tasks and 17 robot embodiments, while EmbodiedSWE-Gen diversifies their solutions into demonstrations that improve robot policy learning and generalization.
TL;DR: KinDER is a physical reasoning benchmark with 25 procedurally generated environments and standardized baselines spanning imitation learning, reinforcement learning, foundation-model planning, diffusion planning, and task and motion planning.
TL;DR: We introduce PlayWorld, a system for training video-based robot world models using autonomous self-play rather than human demonstrations, generating high-quality physically consistent predictions for contact-rich interactions that enable reinforcement learning in simulation with 65% improvements in real-world policy success rates.
Paper Website Code Checkpoints
TL;DR: We introduce Language-Action Pre-training (LAP), which represents robot actions as natural language tokens to enable vision-language-action models to transfer zero-shot to new robot embodiments, achieving over 50% zero-shot success on novel robots—approximately twice the performance of prior methods.
TL;DR: We survey video generation models in robotics, covering current applications, core research challenges, and promising future directions for using generative video models to support robot prediction, planning, and interaction.
Best Paper Award, CoRL 2025 Eval&Deploy workshop
TL;DR: We present a framework that augments real-world evaluations with simulation evaluations to provide stronger inferences on real-world policy performance that could otherwise only be obtained by scaling up real-world evaluations.
TL;DR: We introduce VLM2VLA, a VLA model training paradigm that represents low-level robot actions in natural language to better align the robot fine-tuning data with the base VLM’s representation space. VLM2VLA yields a policy with strong VQA performance and zero-shot generalization to new scenarios.
Best Paper Award, RSS 2025 SWoMo and SemRob workshops
TL;DR: We introduce World Models for Active Perception (WoMAP), a scalable recipe for training open-vocabulary object localization policies that are grounded in the physical world.
Spotlight, RSS 2025 CRLH workshop
Paper Talk Video Website Code Slides
TL;DR: We introduce Factored Scaling Curves (FSC), which model how policy performance scales with data for different environmental factors and can be extrapolated to guide principled data collection.
TL;DR: We propose DROC that can respond effectively to online human language corrections, distill generalizable knowledge from corrections, and retrieve usable knowledge for future tasks.
TL;DR: We show how to leverage LLMs to generate constraints that can indicate misalignment during execution, and use VLMs to detect constraint violations continuously.
Best Presentation Award
TL;DR: We propose using time series reconstruction to process serial wind power data and achieve state-of-the-art wind power prediction accuracy.