FineART

FineART (Fine-Grained Annotated Robotic Trajectory) is a bimanual robot manipulation dataset co-released by Scale AI and Hugging Face LeRobot — Scale AI's first open-source robotics dataset and model. Scale Labs dated the release 2026-10-29, and this week it is climbing the Hugging Face dataset trending list (lerobot/FineART, 840k+ downloads).

Scale:

• 40,543 teleoperated episodes, 1,718 hours, 185.5M frames;
• 533,913 timestamped natural-language subtask annotations (~13 per episode);
• 151 tasks — ~77% pick-and-place/sorting, 23% contact-rich long tail (textiles, cables, tool use, precision insertion);
• Collected on a Trossen Stationary AI (ALOHA-style) bimanual robot with 4 cameras; LeRobot dataset v3.0 format; CC BY 4.0.

Alongside it ships FineART-VLA, an extension of Physical Intelligence's π0.5 combining a System-2 subtask predictor and System-1 continuous-action controller in one policy.

What problem it solves

Most open robot datasets provide one task-level instruction per episode ("unpack the box") or annotate subtasks for only part of their hours, which limits hierarchical "plan the next step, then act" policies. FineART densely annotates 100% of its hours, claiming to be the largest subtask-annotated open robot dataset.

Highlights

• Two-level annotation: task-level plus temporally segmented subtasks, with execution-quality/success metadata;
• Three-stage pipeline: separate teams collect, annotate and validate;
• FineART-VLA (self-reported): mid-training lifts spatial instruction following from 32% to 100%; with human step-by-step subtask guidance, unseen long-horizon success rises from 16% to 76%; matches baseline on a new robot with ~10× less fine-tuning data.

Compared with peers

Versus instruction-per-episode datasets like Open X-Embodiment or DROID, FineART is better suited to System-1/System-2 hierarchical VLAs, but covers a single hardware platform (ALOHA-style bimanual).

Who it's for

• Teams training/evaluating VLAs and hierarchical robot policies;
• Labs using ALOHA / Trossen bimanual rigs;
• Embodied-AI developers studying the language-subtask → action interface.

Getting started

1. pip install lerobot (a version supporting the language_persistent / language_events columns);
2. Load: from lerobot.datasets.lerobot_dataset import LeRobotDataset; ds = LeRobotDataset("lerobot/FineART");
3. Use lerobot-annotate to segment subtasks in your own LeRobot dataset;
4. Fine-tune from lerobot/FineART-VLA-subtask-midtrained with lerobot-train.

Notes

At ~1,718 hours of video, a full download needs substantial storage and bandwidth; stream task subsets first.