Does SmolVLA run on Franka Research 3?
Capabilities
| Capability | State | Notes | Evidence |
|---|---|---|---|
| Train / fine-tune | works | Data recorded on the FR3 in LeRobotDataset format and trained with LeRobot's training scripts. | ev-lerobot-issue-1567-crisp, ev-crisp-paper-deploy, ev-phail-card-smolvla |
| Real-robot inference | works with caveats | ev-crisp-paper-deploy, ev-crisp-paper-fr3, ev-phail-card-smolvla | |
| Simulation inference | no verified evidence | — | |
| Teleop + record | no verified evidence | — |
Caveats
- Community-reported only (CRISP paper authors); not documented upstream or maintainer-tested.
- Runs through the third-party CRISP ROS 2 controllers and crisp_gym interface; LeRobot v0.6.1 has no native FR3 robot type.
- SmolVLA ran at approximately 10 Hz and outputs end-effector target poses tracked by CRISP's Cartesian impedance controllers; the paper reports no task success rates.
Evidence 4 sources; the cell shows the highest tier
- T3 · community reportedpapersource 2026-02-06verified 2026-10-25 (today)
CRISP paper, Fig. 5: Diffusion Policy and SmolVLA trained on CRISP_GYM-recorded data are deployed on hardware; SmolVLA at ~10 Hz.
“Using the same recorded data in the CRISP_GYM interface, different learned policies can be successfully deployed in hardware. Left: Diffusion Policy [4] operating at approximately 30 Hz. Right: SmolVLA [27] operating at approximately 10 Hz.”
https://arxiv.org/abs/2509.06819v2 - T3 · community reportedpapersource 2026-02-06verified 2026-10-25 (today)
CRISP paper, Sec. IV: the only hardware platform is the Franka Robotics FR3 (other arms are simulation only).
“The controllers have been validated on hardware using the Franka Robotics FR3 and in simulation with the Kuka IIWA14 and Kinova Gen3, as shown in Figure 3.”
https://arxiv.org/abs/2509.06819v2 - T3 · community reportedgithub-issuesource 2025-08-14verified 2026-10-25 (today)
CRISP author (LeRobot contributor) reports training VLAs including SmolVLA with LeRobot on a Franka Robotics FR3, recording LeRobotDataset data and deploying with their own script.
“At our lab we have been training VLAs (also tried smolVLA) using lerobot and a Franka Robotics FR3.”
https://github.com/huggingface/lerobot/issues/1567#issuecomment-3188267089 - T3 · community reportedcommunity-postsource 2026-05-06verified 2026-10-25 (today)
PhAIL v1.0 (anonymous benchmark under double-blind review): ~115 SmolVLA evaluation rollouts on a Franka Research 3 with Robotiq 2F-85 (bin-to-bin pick-and-place), SmolVLA fine-tuned on the 449-episode FR3 corpus with its default recipe; no per-policy success rates on the card.
“Hugging Face SmolVLA | `inference` | ~115 | Default fine-tuning recipe”
https://huggingface.co/datasets/phail-anon/phail-v1.0/blob/5bc590213c5d2615eefeb47bc23a700f8c57022a/README.md
SmolVLA inference speed by hardware robot-independent · inference cost only
| Compute | Runtime | Precision | Cams | p50 / p95 ms | Inference | Platform | Source |
|---|---|---|---|---|---|---|---|
| Jetson Orin Nano 8GB (Super) | TensorRT | fp16 | 2 @640x480 | 107.55 / 108.48 | 9.3 Hz | jetpack 7.2.1 · power_mode MAXN_SUPER, clocks pinned with jetson_clocks (GPU 1020 MHz, CPU 1728 MHz) · tensorrt 10.16.2.10conditionsboard: Jetson Orin Nano Engineering Reference Developer Kit Super (8 GB) · l4t: R39.2.1 · os: Ubuntu 24.04.4 LTS · python: 3.12.3 · desktop: running Inference cost only; not task success. 300 s run on synthetic 640x480 observations resized by the exported policy to 512x512; 10 denoising steps; latency is runtime-only over a pre-materialized observation ring. Whole-board RAM in use 1.78 GB (unified memory, desktop running). Actions were gated against the stock policy (cosine >= 0.99999) before timing. | T3 · community reported |
| Jetson Orin Nano 8GB (Super) | ONNX (split) | fp16 | 2 | 189.89 / 190.93 | 5.25 Hz | jetpack 7.2 (L4T R39.2.x, per the repo host-setup doc at this commit) · power_mode MAXN_SUPER, pinned clocks · cuda 13.2 · tensorrt 10.16.2 (system, via ONNX Runtime TensorRT EP)conditionsboard: Jetson Orin Nano Super 8 GB developer kit · os: Ubuntu 24.04 · onnxruntime-gpu: 1.24 Inference cost only; not task success. Deterministic synthetic observations, pinned MAXN_SUPER clocks. Configured EP priority 7 graphs TensorRT / 2 CPU. System RAM 2385 MB in use. | T3 · community reported |
| Jetson Orin Nano 8GB (Super) | PyTorch eager | fp32 | 2 | 1167.93 / 1176.65 | 0.86 Hz | jetpack 7.2 (L4T R39.2.x, per the repo host-setup doc at this commit) · power_mode MAXN_SUPER, pinned clocks · cuda 13.2 · lerobot 0.5.1conditionsboard: Jetson Orin Nano Super 8 GB developer kit · os: Ubuntu 24.04 Inference cost only; not task success. Deterministic synthetic observations, pinned MAXN_SUPER clocks. Stock LeRobot SmolVLAPolicy in PyTorch FP32 (LeRobot 0.5.1 env, JetPack-matched torch). RSS peak 3826 MB, system RAM 4126 MB in use. | T3 · community reported |
Embed this cell
Badge for READMEs and model cards. It always shows the current state.
[](https://caniembody.pages.dev/compat/franka-fr3/smolvla)<iframe src="https://caniembody.pages.dev/embed/franka-fr3/smolvla" width="380" height="132" style="border:0" title="SmolVLA on Franka Research 3 | CanIEmbody"></iframe>History
- 2026-10-25 fecccd1 Initial release: robot policy compatibility matrix · Danilo Znamerovszkij