Previous work, August 2026
The first complete system: 150 teleoperated demonstrations on the Jetson Thor rig with three cameras, and a π0.5 policy that follows source-to-destination commands.
The policy keeps π0.5's vision–language backbone frozen and trains only its action expert. A person scored 79 of its rollouts over 15 commands; the numbers below come from that evaluation. The newer π0.5 and FLUX runs, on a workstation with five cameras, are the latest work.
How well it worked
- 25%completed the whole pick-and-place20 of 79 rollouts
- 8%acted on the wrong vial or slot6 of 79 rollouts
- 54%missed the first grasp of the vial43 of 79 rollouts
- 44%completed the move once the first grasp held16 of 36 rollouts
Everything from August
- The evaluationWhere the takes were lost, command by command, with every scored trial on video.
- The datasetsVersions 1 to 5, how the demonstrations were recorded, and why the policy uses no depth.
- The reportAbstract, PDF and citation.
- The 150 demonstrationsOn Hugging Face, in the LeRobot format.
The policies tried in August
Only π0.5 was deployed and evaluated; the others were set up on the same rig and not carried through.
| Policy | What was trained | Data | Steps | Training time | Peak GPU memory | On the robot | Source |
|---|---|---|---|---|---|---|---|
| The August policies (Thor rig, three cameras) | |||||||
π0.5, frozen backbone π0.5 | Action expert only; vision-language backbone frozen | vial-sort-v5, 150 demos, 3 cameras | 30k (12k deployed) | not logged | not logged | 20 of 79 rollouts complete, scored by a person. | report |
π0 π0 | Action expert, full fine-tune vs LoRA; backbone frozen | vial-sort-v1, 100 demos | 30k | about 5–6 h each | same for both | Same training loss, but LoRA sometimes missed the grasp; full fine-tune grasped reliably. Only one fixed command was tried. | project notes (June 2026), report |
GR00T N1.5 GR00T N1.5 | Fine-tune, batch 16 | vial-sort-v4 | 20k | 3 h 58 min | 54.7 GB | Exploratory; not carried through to deployment. | project notes (July 2026), report |
ACT ACT | – | – | – | not logged | not logged | No language input; a motion baseline only. | report |