Any eval for this port?
Is there an evaluation run for this build, on LIBERO or a real arm, to confirm it behaves like the upstream pi0.5 checkpoint? Even one small table would help people trust the port.
Thanks for asking. Yes — this checkpoint has been evaluated, on both a real arm and in simulation:
- Real SO-101, tabletop banana pick (banana placed by hand at conventional positions): > 80% success.
- Simulation, same task (fully randomized over all reachable positions, extensive trials): ~60%.
- Simulation + RL: up to 84% on the same randomized protocol. Checkpoints for these experiments, including the RL one, will be open-sourced progressively.
The simulation number is lower because it sweeps the entire reachable workspace while the real-arm figure uses the ordinary placement band that the demonstrations cover densely — the harder distribution, and the more conservative number.
The IB-Robot training guide has been updated with the PI05 fine-tuning and deployment steps (dataset preparation, training commands, checkpoint -> bundle handoff and registration), so the pipeline behind these numbers can be reproduced.