PhD Student @ TEALab, Tsinghua Univ.; Past Visiting Scholar @ TML, Stanford University

• Pinned

  

  

• 

  Robot learning policies are data-hungry and still fail on objects they weren't trained on. We propose AffordGen, from ONE human demo, we synthesize thousands of trajectories that generalize to unseen objects and even unseen categories, with the help of affordance correspondence.

  

  00:00

• 

  Checkout our new realworld RL framework that keeps recovery potential in mind! We define 'critial failure' during realworld learning and trying to avoid them during the rollout.

  

  There's a dilemma: We want robot agents that adapt online, but we can't afford the cleanup costs when they explore. 💥 Here we are. Introducing FARL: Failure-Aware Offline-to-Online RL. We enable robots to fine-tune in the real world while actively avoiding disasters.

  

  00:00

• 

  Congrats, Kun! Extending the successful real-world RL experiences from Uni-O4 (lei-kun.github.io/uni-o4/) to robot manipulation tasks is just amazing!

  

  Introducing RL-100: Performant Robotic Manipulation with Real-World Reinforcement Learning. lei-kun.github.io/RL-100/ 7 real robot tasks, 900/900 successes. Up to 250 consecutive trials in one task, running 2 hours nonstop without failure. High success rate against physical

  

  00:00

• 

  It was very emotional moment to watch Arya, the second Toddlerbot, training with the RTR hardware.