Open robot dataset with 40K+ demonstrations and a step-by-step reasoning model
The Hook
“(No spoken dialogue — ambient electronic synth score)”
Product Reveal
“We're introducing FineART, Scale's first open-source robotics dataset. FineART is the world's largest subtask-annotated data. With Hugging Face's LeRobot team, we're releasing FineART-VLA, the first open-source System 1-System 2 policy.”
Feature Teaser
“Our dataset features over 40,000 episodes, 100% subtask-annotated. With more than half a million subtasks, we enable learning, step by step, at scale.”
Feature Teaser
“FineART-VLA combines a high-level planner with a low-level motor executor into a single unified architecture. FineART-VLA is the first dual-system VLA natively integrated in LeRobot.”
Feature Teaser
“FineART-VLA is trained on three objectives: tokenized actions, tokenized text, which are the subtasks, and continuous actions. The model trained on our annotations has two modes of inference: hierarchical inference and flat mode. Hierarchical inference allows the model to predict its own subgoal before acting in the real world. The dataset, including the annotation and model weights and training code, are available on Hugging Face.”
Feature Teaser
“We then tested our policy in the real world, benchmarking across embodiments against baselines with and without mid-training. Spanning 3,400 rollouts, it is one of the largest evaluations on record.”
Feature Teaser
“Mid-training on FineART's subtask annotation raised the success rate of our policy on spatial-instruction tasks by 68 percentage points. For instance, it can recognize the difference between left, right, or center. It also boosts long-horizon performance, raising success rate of our prepare-breakfast task by 60 percentage points. We also see zero-shot transfer from the model on a completely new task and on a new target embodiment, which has a completely different morphology.”
Call to Action
“Are you ready to build? Annotate your dataset with LeRobot Annotate, fine-tune FineART-VLA, and get started. Everything is open source. At Scale, we are pushing the frontier of robotics research with high-quality, high-diversity data. We work on developing new multimodal capabilities with leading labs and organizations like Hugging Face. FineART is our first open-source robotics dataset and model, built to help the community move faster. This is just a canvas, and we can't wait to see what you build with it. After all, robotics is a fine art.”
Synthesizable in Remotion · key components of the open-source-robotics-launch template.
A responsive 4x4 matrix grid displaying 16 synchronized video rollouts simultaneously.
Map an array of 16 video sources within a CSS grid container (grid-template-columns: repeat(4, 1fr)) using Remotion Video components with shared startFrom offset.
Technical neural network diagram highlighting active input tokens, VLM backbone blocks, and action expert outputs.
Use SVG containers styled with CSS flexbox, driving node opacity and token highlight transitions via interpolate(frame, [0, 30], [0, 1]).
Three synchronized vertical video feeds side-by-side highlighting micro-manipulation tasks.
Create 3 flex child containers with width 33.33% and overflow hidden, positioning Video elements with object-fit cover.
Purpose. Establish high-trust partnership between Scale AI and Hugging Face.
Execution. Lab establishing shot overlaid with joint logos.
Purpose. Present the 40k+ episode milestone and subtask granularity.
Execution. On-camera researcher narrative intercut with annotation tool screens.
Purpose. Detail dual System 1 / System 2 architecture and loss objectives.
Execution. High-contrast animated architectural schematic.
Purpose. Validate claims using extensive benchmarks and spatial manipulation tests.
Execution. 16-screen rollout grid followed by top-down robotic pick-and-place footage.
Purpose. Invite developer community to build using open-source weights.
Execution. Leadership wrap-up speech concluding on hardware logo lockup.