Is FineART-VLA state of the art?Pending benchmarks.Announced Oct 29, 2026; no independent results yet.
An open vision-language-action (VLA) robot model from Scale AI and Hugging Face's LeRobot team that plans before it acts: a System 2 part predicts the next subtask in words and a System 1 part turns it into arm actions. Fine-tuned from Physical Intelligence's pi0.5 on FineART, Scale's dataset of 40,543 two-arm robot episodes with 533,913 subtask labels. About 4.1B parameters; weights on Hugging Face under the Gemma license.
The details so far
- Lab
- Scale AI (Scale Labs) with Hugging Face's LeRobot team; published in the lerobot Hugging Face org
- Announced
- Oct 29, 2026
- API
- Not yet
- Open weights
- Published · Hugging Face, about 9.4 GB (F32 and BF16 safetensors) under the Gemma license inherited from PaliGemma/pi0.5; LeRobot code is Apache-2.0
- Parameters
- About 4.1B (4,143,404,816 in the safetensors): a Gemma 2B vision-language backbone plus a Gemma 300M action expert
- Type
- Vision-language-action robot policy, not a chat model: camera frames, a language instruction and robot state in, actions out
- Design
- System 2 predicts subtasks in natural language; System 1 turns the current subtask into actions (50-step action chunks, 10 flow-matching steps)
- Base model
- lerobot/pi05_base, LeRobot's port of Physical Intelligence's pi0.5 (PaliGemma-based)
- Robot
- Stationary two-arm ALOHA-style setup: three cameras (high, left wrist, right wrist) at 224×224, 14-dimensional state and action
- Training data
- FineART dataset (CC-BY-4.0): 40,543 episodes, 1,718 hours, 151 tasks, 533,913 subtask annotations (about 13 per episode)
- Code
- Training and fine-tuning through LeRobot (lerobot-train; policy fineart_vla); lerobot-annotate labels subtasks on other LeRobot datasets
- Paper
- arXiv 2609.36416, first posted Sep 29, 2026 (UTC); the weights followed on Oct 29, 2026
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: FineART is the largest open robot dataset with subtask labels.
Lab claim: across 3,400 real-world evaluations, subtask training lifts spatial-instruction success from 32.0% to 100.0% and, with step-by-step guidance, unseen long-horizon task success from 16.0% to 76.0%.
Lab claim: on a new robot (YAM arms) it matches the baseline with 10x less fine-tuning data.
Sources
LABScale Labs blog: FineART, Scale AI's first open-source robotics dataset and modellabs.scale.com/blog/fineart · Oct 29, 2026LABScale Labs announcement on Xx.com/ScaleAILabs/status/2115837725719048535 · Oct 29, 2026REPOlerobot/FineART-VLA-subtask-midtrained on Hugging Facehuggingface.co/lerobot/FineART-VLA-subtask-midtrainedREPOlerobot/FineART dataset on Hugging Facehuggingface.co/datasets/lerobot/FineARTLABLeRobot docs: FineART-VLAhuggingface.co/docs/lerobot/main/fineart_vlaPAPERFineART paper (arXiv 2609.36416)arxiv.org/abs/2609.36416Tracking continues until an independent board lists it.