Does MolmoAct2 run on SO-100?
Capabilities
| Capability | State | Notes | Evidence |
|---|---|---|---|
| Train / fine-tune | works with caveats | Fine-tuning from the SO-100/101 checkpoint needs the same joint frame transform. | ev-lerobot-molmoact2-calib, ev-molmoact2-so100-101-card |
| Real-robot inference | works with caveats | ev-lerobot-molmoact2-doc, ev-lerobot-molmoact2-calib, ev-molmoact2-so100-101-card | |
| Simulation inference | no verified evidence | — | |
| Teleop + record | no verified evidence | — |
Caveats
- The MolmoAct2-SO100_101 weights expect the pre-LeRobot-0.5.0 joint calibration convention: lerobot/MolmoAct2-SO100_101-LeRobot includes the frame correction, but self-converted or fine-tuned checkpoints must set joint_signs/joint_offsets or the arm may move in the wrong direction.
- Camera keys must be renamed to cam0/cam1 (--rename_map) to match the checkpoint.
Evidence 3 sources; the cell shows the highest tier
- T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
LeRobot v0.6.1 MolmoAct2 guide: zero-shot rollout of lerobot/MolmoAct2-SO100_101-LeRobot on so100_follower (top+side cameras renamed to cam0/cam1).
“Example with an SO-100 robot using top and side cameras:”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/molmoact2.mdx - T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
LeRobot v0.6.1 MolmoAct2 guide warns that the SO-100/101 checkpoint uses the pre-0.5.0 joint calibration convention; the converted LeRobot checkpoint includes the frame correction, self-converted/fine-tuned ones need joint_signs/joint_offsets.
“The MolmoAct2-SO100_101 checkpoint was trained on data that uses a different joint calibration convention than LeRobot >= 0.5.0. Without a frame correction, the arm may move in the wrong direction.”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/molmoact2.mdx - T1 · upstream documentedmodel-cardsource 2026-05-23verified 2026-10-24 (1d ago)
Ai2's official MolmoAct2-SO100_101 model card: fine-tuned on an SO-100/101 data mixture, intended for SO-100/101 inference and further fine-tuning.
“This checkpoint is fine-tuned on the SO-100/101 mixture with absolute joint-pose control and annotated language instructions. It is intended for both further fine-tuning and SO-100/101 policy inference.”
https://huggingface.co/allenai/MolmoAct2-SO100_101/blob/152569fe57914d97be91055800035f54e250d009/README.md
MolmoAct2 inference speed by hardware robot-independent · inference cost only
| Compute | Runtime | Precision | Cams | p50 / p95 ms | Inference | Mem | Platform | Source |
|---|---|---|---|---|---|---|---|---|
| H100 80GB | PyTorch eager | bf16 | 2 | — | — | 13 GB | lerobot 0.6.1conditionsos: Ubuntu 22.04 (repo tested with) · memory_sampling: nvidia-smi peak Inference memory only (12.1 GiB peak GPU memory = 13.0 GB); no latency reported. LIBERO with two RGB cameras, chunk_size=10, 8 flow timesteps. Not task success. | T3 · community reported |
Embed this cell
Badge for READMEs and model cards. It always shows the current state.
[](https://caniembody.pages.dev/compat/so100/molmoact2)<iframe src="https://caniembody.pages.dev/embed/so100/molmoact2" width="380" height="132" style="border:0" title="MolmoAct2 on SO-100 | CanIEmbody"></iframe>History
- 2026-10-25 fecccd1 Initial release: robot policy compatibility matrix · Danilo Znamerovszkij