IEEE Spectrum· Jun 30, 2025
New Humanoid Robot from Fourier Intelligence
Victor Suen of Fourier Intelligence introduces the company’s GR-1 humanoid robot and discusses its origins, intended uses, technical focus, and open platform.
Ep#3 Sim-to-Real RL for Vision-Based Dexterous Manipulation on Humanoids
Toru, a UC Berkeley PhD student working with Jitendra Malik, discusses vision-based sim-to-real reinforcement learning for humanoid manipulation with RoboPapers hosts.
In this talk
- 0:55
The robot and its visual inputs
Toru presents a system tested on a Fourier GR-1 humanoid with two Shadow Finger hands. It combines egocentric RGB-D depth with a third-person RGB-D view that helps locate objects across a wider workspace. The policy uses segmented depth and object key points, not RGB pixels; users select the object through a UI backed by Segment Anything.
- 9:53
Why move from demonstrations to reinforcement learning
Toru describes earlier work collecting teleoperated demonstrations and says hand teleoperation can be difficult because of retargeting, latency, operator training, and inconsistent human motions. Data can also stop working when conditions such as lighting or camera placement change. These challenges led Toru to explore simulation-trained reinforcement learning, first with a two-hand twisting task and later with broader humanoid manipulation.
- 21:45
Rewards, simulation tuning, and training
Toru structures rewards around contact goals, such as where fingers should touch an object, and object goals, such as where it should move. To tune the robot model, the team compares simulated and real response curves across a broad test trajectory and selects the URDF parameters with the lowest mean squared error. A brief human-guided hand-pose initialization in simulation narrows exploration; grasping policies trained on four L4 GPUs took under 24 hours, while handover took two to three days.
- 32:12
Results and remaining limits
The system demonstrates grasping, lifting, and handover, including grasping objects outside the training distribution with one shared grasping policy. Toru notes that the six-degree-of-freedom hands limit the tasks they can perform, and reported success rates plateau around 50 to 60 percent for some tasks. Future improvements could address dynamics mismatch, perception, and reliance on a fixed third-person camera; active vision is one proposed direction.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.