Our research sits at the intersection of robotics, machine learning, and embodied AI, with a focus on foundation models, efficient learning, and human-centered intelligence.
We aim to build physical intelligent agents that can perceive, reason, and act in the real world.
We develop efficient foundation models that unify perception, reasoning, understanding, and action for embodied agents. This includes vision-language-action (VLA) models, world models, and generative AI methods that let agents understand complex scenes, reason about long-horizon goals, and act reliably under real-world limits on compute, latency, and data.
We build foundation-model-based systems that let mobile robots and manipulators operate in unstructured environments. Our work spans visual navigation, dexterous manipulation, multi-modal scene understanding, and decision-making and control, turning strong semantic and geometric representations into robust closed-loop robot behavior.
We study how robots can adapt quickly, interact, and learn from humans with limited supervision. This includes data-efficient fine-tuning, reinforcement learning, imitation learning, and human-in-the-loop adaptation, so robots can personalize their behavior, improve through interaction, and generalize to new tasks, users, and environments.
Recent work by Prof. Huang and collaborators, the foundation the PARIS Lab builds on. All clips are real rollouts of our models, in the real world and in simulation.
Learning Effective Latent Memory for Long-Horizon Robotic Manipulation
Most VLA policies act from the current frame alone, but long-horizon manipulation depends on what the robot saw earlier. Mem4Act distills past observations into a compact, fixed-size latent memory trained explicitly for control, reaching state-of-the-art success on memory-dependent tasks in simulation (RoboMME, RMBench) and on a real dual-arm robot.
Instruction: Cover the red, green, and blue blocks with cups from left to right, then uncover them in the order red, green, blue.
A Think-in-Control Vision-Language-Action Model for Robot Navigation in Dynamic Environments
Large VLA models think slowly, but robots must act now. TIC-VLA is a latency-aware framework that explicitly models delayed semantic reasoning during action generation, so a quadruped can follow language instructions through dynamic, crowded indoor and outdoor spaces in real time.