Does Multitask DiT run on Franka Research 3?

works with caveats

T1 · upstream documentedlast verified 2026-10-25 · todayLeRobot

Capabilities

Capability	State	Notes	Evidence
Train / fine-tune	works	LBMs are pretrained on a multi-source mixture (TRI's own real data, lbm_real, weighted 0.5) and finetuned per task (30k steps for real tasks).	ev-lbm-paper-finetune, ev-lbm-paper-platform
Real-robot inference	works	Real-world evaluation on nine bimanual FR3 stations; the policy runs at 10 Hz.	ev-lbm-paper-platform, ev-lbm-paper-hardware, ev-lbm-paper-dit
Simulation inference	no verified evidence		—
Teleop + record	no verified evidence		—

Caveats

Evidence is TRI's own Large Behavior Model (DiT) stack, not LeRobot's multi_task_dit reimplementation; LeRobot has no FR3 robot type and no FR3 run through LeRobot is documented.
TRI's platform is bimanual: two FR3 arms per station with custom compliant fingers (Schunk grippers on the evaluated platform) and 4 wrist cameras.

Evidence 5 sources; the cell shows the highest tier

T1 · upstream documentedpapersource 2025-07-07verified 2026-10-25 (today)
TRI LBM paper, Sec. 4.3.1: the real-world platform is tabletop bimanual manipulation with two Franka FR3 arms.
“We focus here on tabletop bimanual manipulation using two Franka FR3 robot arms with parallel gripper and TRI’s finray-style fingers”
https://arxiv.org/abs/2507.05331v1
T1 · upstream documentedpapersource 2025-07-07verified 2026-10-25 (today)
TRI LBM paper, Sec. 7.1: every robot station has two Franka Research 3 arms.
“All stations contain a tabletop marked with grid lines (useful for correlating object placement on the real robot stations), two Franka Research 3 arms, and two parallel jaw grippers with custom compliant fingers”
https://arxiv.org/abs/2507.05331v1
T1 · upstream documentedpapersource 2025-07-07verified 2026-10-25 (today)
TRI LBM paper, Fig. 9: the LBM policy is a language-, vision- and proprioception-conditioned Diffusion Transformer run at 10 Hz on deployment.
“conditioned on language, vision and proprioception and which outputs 20-dimensional actions over 16 timesteps”
https://arxiv.org/abs/2507.05331v1
T1 · upstream documentedpapersource 2025-07-07verified 2026-10-25 (today)
TRI LBM paper: LBMs are finetuned per task on real (and simulated) data before evaluation.
“we finetune for 30k steps for real tasks and for 10k steps for simulated tasks with a global batch size of 320 and a reduced learning rate of 2e-5.”
https://arxiv.org/abs/2507.05331v1
T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-25 (today)lerobot 0.6.1
LeRobot v0.6.1 docs: LeRobot's Multitask DiT diffusion implementation is based on TRI's LBM paper, which ties this grid policy to the LBM results.
“The original diffusion implementation here is based on the work described in [TRI's LBM paper](https://arxiv.org/abs/2507.05331)”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/multi_task_dit.mdx

Embed this cell

Badge for READMEs and model cards. It always shows the current state.

[![Multitask DiT on Franka Research 3: works with caveats](https://caniembody.pages.dev/badge/franka-fr3/multitask-dit.svg)](https://caniembody.pages.dev/compat/franka-fr3/multitask-dit)

<iframe src="https://caniembody.pages.dev/embed/franka-fr3/multitask-dit" width="380" height="132" style="border:0" title="Multitask DiT on Franka Research 3 | CanIEmbody"></iframe>

History

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