Portfolio item number 1
Short description of portfolio item number 1
Posts by Collection
portfolio
Portfolio item number 2
Short description of portfolio item number 2
publications
DittoGym: Learning to Control Soft Shape-Shifting Robots
ICLR 2024 [Project]
A modeling of the highly-reconfigurable robot within a MPM-powered simulator, along with effective coarse-to-fine curriculum for high-dimensional RL policy training and a comprehensive reconfigurable robot testbed.
MENTOR: Mixture-of-Experts Network with Task-Oriented Perturbation for Visual Reinforcement Learning
ICML 2025 [Project]
A model-free visual RL algorithm achieving state-of-the-art performances in learning efficiency and performance on various tasks. The policy can even be directly trained on real robot to play tabletop golf!
ParticleFormer: A 3D Point Cloud World Model for Multi-Object, Multi-Material Robotic Manipulation
CoRL 2025 [Project]
A state-of-the-art 3D point cloud world model for complex scene dynamics modeling, enabling accurate dynamics prediction across multi-object and multi-material scenarios, and empowering model-based visuomotor control with novel configurations.
GRaD-Nav++: Vision-Language Model Enabled Visual Drone Navigation with Gaussian Radiance Fields and Differentiable Dynamics
RA-L 2025 [Best Paper Award] [Project]
A lightweight, fully onboard VLA framework for real-time language-guided drone control. Powered by 3D-GS differentiable simulation and an MoE-based policy backbone, it efficiently learns generalizable policies that transfer from simulation to the real world.
MoE-DP: An MoE-Enhanced Diffusion Policy for Robust Long-Horizon Robotic Manipulation with Skill Decomposition and Failure Recovery
ICRA 2026 [Project]
A robust long-horizon robotic manipulation framework integrating Mixture-of-Experts into diffusion policies for skill decomposition, failure recovery and resilient visuomotor control across multi-stage tasks.
Breaking Lock-In: Preserving Steerability under Low-Data VLA Post-Training
arXiv [Project]
A generalization-preserving VLA post-training method that mitigates policy lock-in under low-data post-training, maintaining pretrained visual grounding and enabling steerable instruction following across novel concepts, spatial targets, and task configurations, without requiring extra adaptation data!