Research

  1. EmbodiedSWE: Coding Agents for Long-Horizon Dexterous Robotics
    EmbodiedSWE: Coding Agents for Long-Horizon Dexterous Robotics
    Haoxiang You, Zeyu Shen, Yilang Liu, Zhicheng Zheng, Lihan Zha, Kashu Yamazaki, Mingtong Zhang, Suning Huang, Jiankai Sun, Qianzhong Chen, Lucy He, Haoran Chang, Dhruv Shah, Mac Schwager, Katerina Fragkiadaki, Peter Henderson, Ian Abraham, and Canwen Xu
    under review 2026
    We systematically evaluate frontier coding agents on complex, long-horizon everyday robotics tasks, and propose a data-engine pipeline that distills the agents’ capabilities into a local VLA policy deployable on a real robot.
    Website arXiv Code
  2. Efficient On-policy Visual-RL via Stochastic Decoupled Policy Gradient
    Efficient On-policy Visual-RL via Stochastic Decoupled Policy Gradient
    Haoxiang You, Yilang Liu, Davis Zong, Qian Wang, Teeratham Vitchutripop, Qi Wang, Daniel Rakita, and Ian Abraham
    CoRL 2026
    A lightweight visual RL method that trains visuomotor policies end-to-end within a few hours on a single GPU, with sim-to-real transfer.
    Website arXiv Code
  3. Accelerating Visual-Policy Learning through Parallel Differentiable Simulation
    Accelerating Visual-Policy Learning through Parallel Differentiable Simulation
    Haoxiang You, Yilang Liu, and Ian Abraham
    NeurIPS 2025 Spotlight
    Accelerates visual RL training with differentiable simulation by decoupling rendering from the gradient computation; a humanoid learns to run from raw pixels in 4 hours on a single GPU.
    Website arXiv Code
  4. Sample-Based Hybrid Mode Control: Asymptotically Optimal Switching of Algorithmic and Non-Differentiable Control Modes
    Sample-Based Hybrid Mode Control: Asymptotically Optimal Switching of Algorithmic and Non-Differentiable Control Modes
    Yilang Liu, Haoxiang You, and Ian Abraham
    ICRA 2026
    A sample-based search over different control modes and when to switch between them, enabling agile robot control.
    arXiv
  5. Goal-Driven Reward by Video Diffusion Models for Reinforcement Learning
    Goal-Driven Reward by Video Diffusion Models for Reinforcement Learning
    Qi Wang, Mian Wu, Yuyang Zhang, Mingqi Yuan, Wenyao Zhang, Haoxiang You, Yunbo Wang, Xin Jin, Xiaokang Yang, and Wenjun Zeng
    CVPR 2026
    Replaces hand-designed rewards with pretrained video diffusion models, rewarding agents for following generated goal videos, and even outperforms heavily engineered dense rewards on several manipulation tasks.
    Website arXiv
  6. Is Bellman Equation Enough for Learning Control?
    Is Bellman Equation Enough for Learning Control?
    Haoxiang You, Lekan Molu, and Ian Abraham
    preprint 2025
    Shows that in continuous state spaces the Bellman equation can have exponentially many solutions, only one of them stable, and proposes a positive-definite network that guarantees converging to it.
    arXiv