Does MolmoAct2 run on Franka Emika Panda?

works with caveats

T1 · upstream documentedlast verified 2026-10-24 · 1d agoMolmoAct2 (allenai/molmoact2)

Capabilities

Capability	State	Notes	Evidence
Train / fine-tune	works		ev-molmoact2-paper-droid-ft
Real-robot inference	works	Zero-shot real-world DROID evaluation: five kitchen tasks x three spatial variants, 87.1% (vs. MolmoBot and π0 baselines).	ev-molmoact2-paper-droid-real, ev-molmoact2-droid-server, ev-lerobot-molmoact2-doc, ev-droid-paper
Simulation inference	no verified evidence		—
Teleop + record	no verified evidence		—

Caveats

Evidence is for the DROID setup (Franka, external + wrist camera, 15-step chunks at DROID's 15 Hz). The paper says 'DROID setup' without naming the arm model. The DROID platform is Panda-based, but its software also accepts an FR3.
The upstream repo ships the MolmoAct2-DROID inference server (examples/droid/host_server_droid.py) but not the robot-side client.

Evidence 5 sources; the cell shows the highest tier

T1 · upstream documentedpapersource 2026-05-08verified 2026-10-24 (1d ago)
MolmoAct2 paper, Sec. 6.2 / App. C.2: MolmoAct2-DROID is evaluated zero-shot on a real DROID setup in novel scenes, reaching 87.1%.
“In real-world evaluation on the DROID setup, which is substantially more out-of-distribution given the random camera initialization and entirely novel scenes and objects relative to the DROID dataset (Khazatsky et al., 2024), MolmoAct2 again attains the highest performance, reaching 87.1%”
https://arxiv.org/abs/2605.02881v2
T1 · upstream documentedpapersource 2026-05-08verified 2026-10-24 (1d ago)
MolmoAct2 paper: the MolmoAct2-DROID checkpoint is fine-tuned on a quality-filtered DROID mixture.
“The MolmoAct2-DROID checkpoint is fine-tuned on the filtered DROID mixture.”
https://arxiv.org/abs/2605.02881v2
T1 · upstream documentedsource-codesource 2026-10-02verified 2026-10-24 (1d ago)
Upstream MolmoAct2-DROID HTTP inference server (allenai/MolmoAct2-DROID, norm tag franka_droid). It takes an external + wrist image and an 8-D joint+gripper state.
“MolmoAct2-DROID inference server.”
https://github.com/allenai/molmoact2/blob/6070080a20321b4f498ab30f28e1d09ac465edb7/examples/droid/host_server_droid.py
T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
LeRobot v0.6.1 lists a LeRobot-format MolmoAct2-DROID checkpoint for lerobot-rollout. LeRobot has no Panda robot type, so that route needs a third-party robot integration.
“| MolmoAct2-DROID-LeRobot | `cam0`, `cam1` | External + wrist |”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/molmoact2.mdx
T1 · upstream documentedpapersource 2025-04-22verified 2026-10-24 (1d ago)
The DROID platform and dataset are built on the Franka Emika Panda.
“We chose the Franka Emika Panda 7 DoF robot arm as the base of our setup since it is widely adopted in the robot research community, reliable, relatively affordable and was available at most participating institutions.”
https://arxiv.org/abs/2403.12945v2

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 0.6.1 conditions os: 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.

[![MolmoAct2 on Franka Emika Panda: works with caveats](https://caniembody.pages.dev/badge/franka-panda/molmoact2.svg)](https://caniembody.pages.dev/compat/franka-panda/molmoact2)

<iframe src="https://caniembody.pages.dev/embed/franka-panda/molmoact2" width="380" height="132" style="border:0" title="MolmoAct2 on Franka Emika Panda | CanIEmbody"></iframe>

History

2026-10-25 fecccd1 Initial release: robot policy compatibility matrix · Danilo Znamerovszkij