Jia Vision's GigaBrain-0.5M* VLA evolves post RoboChallenge win, using world model-conditioned RL for robust long-horizon robotics. Achieves zero-fail hours in folding, coffee service, box tasks via human-in-loop iteration. Four-stage training: pretrain, fine-tune, deploy, evolve.
Key Points
- 1.World model predicts states/values
- 2.Human-loop trajectory optimization
- 3.Stable real-world robot operation
Impact Analysis
Pushes embodied AI toward self-evolution; excels in diverse physical tasks.
Technical Details
Iterative RL with condition inputs; refines via screened real trajectories.