This is Intelligent Robotics and Interactive Systems (IRIS) Lab! Our research focuses include

                   


Recent Updates

Oct 15, 2024

🔥🔥 “Skills from YouTube, No Prep!” 🔥🔥 Can robots learn skills from YouTube without complex video processing? Our "Language-Model-Driven Bi-level Method” makes it possible! By chaining VLM & LLM in a bi-level framework, we use the “chain rule” to guide reward learning directly from video demos. 🚀Check out our RL agents mastering skills from their biological counterparts!🚀


Check out the preprint. Here is a long demo:




Aug 24, 2024

🚀 Can a robotic hand master dexterous manipulation in just 2 minutes? YES! 🎉 Excited to share our recent work “ContactSDF”, a physics-inspired representation using signed distance functions (SDFs) for contact-rich manipulation, from geometry to MPC. 🔥 Watch a full, uncut video of Allegro hand learning from scratch below! We are pushing the boundaries of “fast” learning and planning in dexterous manipulation.


Check out the webpage, preprint, and code. Here is a long demo:




Aug 19, 2024

Can model-based planning and control rival or even surpass reinforcement learning in challenging dexterous manipulation tasks? YES! The key lies in our new "effective yet optimization-friendly multi-contact model."

🔥 Thrilled to unveil our work: "Complementarity-Free Multi-Contact Modeling and Optimization," which consistently achieves state-of-the-art results across different challenging dexterous manipulation tasks, including fingertip 3D in-air manipulation, TriFinger in-hand manipulation, and Allegro hand on-palm manipulation, all with different objects. Check out the demo below!

Our method sets a new benchmark in dexterous manipulation:
  • 🎯 A 96.5% success rate across all tasks
  • ⚙️ High manipulation accuracy: 11° reorientation error & 7.8 mm position error
  • 🚀 Model predictive control running at 50-100 Hz for all tasks

Check out our preprint, and try out our code (fun guaranteed). Here is a long demo:




July 9 2024

🤖 Robots may be good at inferring a task reward from human feedback, but how about inferring safety boundaries from human feedback? In many cases such as robot feeding and liquid pouring, specifying user-comfortable safety constraints is more challenging than rewards. Our recent work, led by my PhD student Zhixian Xie, shows that this is possible, and can actually be very human-effort efficient! Our method is called Safe MPC Alignment (submitted to T-RO), enabling a robot to learn its control safety constraints with only a small handful of human online corrections!

Importantly, the Safe MPC Alignment is certifiable: providing an upper bound on the total number of human feedback in the case of successful learning of safety constraints, or declaring the misspecification of the hypothesis space, i.e., the true implicit safety constraint cannot be found within the specified hypothesis space.

Check out the project website, preprint, and a breaf introduction vide below.