Ep#30: R2S2: Unleashing Humanoid Reaching Potential via Real-world-Ready Skill Space

In this talk

1. 1:27

   Why humanoids need a larger workspace

   Lee says humanoid robots should reach across the 3D workspace people use, including the floor and high surfaces. R2S2 targets the sim-to-real gap that makes multitask reinforcement learning difficult to deploy reliably.

2. 7:40

   Building and combining primitive skills

   R2S2 starts with local motion, body pose adjustment, and hand reaching. The team uses priors such as gait and leg symmetry to stabilize individual skills, then combines imitation learning with reinforcement learning to teach transitions and coordination while preserving the primitive skills.

3. 22:44

   Using the skill space for tasks and teleoperation

   A high-level policy can select latent actions for tasks such as touching target points or picking up a box. For teleoperation, users send hand targets from a VR headset and control lower-body actions with buttons; the same skill prior supports H1 and G1 robots.

4. 50:36

   Evaluating transfer and discussing next steps

   The team measures reachable space, task success, and distance from the end effector to target points; real-world distance measurements use motion capture. Lee says removing height or torso-pitch control shrinks the workspace, while future work could add skills such as stair climbing and improve flexibility and vision-based autonomy.