AutoRT
Uses foundation models to orchestrate fleets of robots collecting their own training data in the wild.
About
AutoRT (January 2024, arXiv 2401.12963) attacks the scarcest resource in robotics: diverse real-world data. Instead of a lab collecting demos, a fleet of 20+ robots roams office buildings while foundation models run the show — a VLM reads the scene, an LLM invents novel instructions ('pick up the fallen napkin'), and the robots execute via teleoperation or existing policies like RT-1 and RT-2. The run collected 77,000 real episodes with minimal human supervision, and the authors show the in-the-wild data is measurably more diverse than lab datasets and improves RT-1 when co-fine-tuned. It is a data-collection system rather than a single downloadable model, and DeepMind released no public code or weights — so treat it as influential infrastructure research, not something you can run.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper (arXiv 2401.12963): 77k episodes across 20+ robots; in-the-wild data improves RT-1 co-fine-tuning (author-reported)
- No independent replication found
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER metrics, metrics-v1). Evidence: Google DeepMind paper (arXiv 2401.12963) with a 77k-episode multi-building deployment and demonstrated gains on RT-1 co-fine-tuning. Strengths: influential data-scaling recipe, honest safety/autonomy tradeoff analysis. Docked for: (1) no public code or weights — it is a system, not an artifact; (2) all results author-reported. License: none — no public release.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 4.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · DeepMind paper arXiv 2401.12963 (Jan 2024); 77k real episodes collected by 20+ robots across multiple buildings
- · No public code or weights released — a research system, not a downloadable artifact
- · In-the-wild data shown to improve RT-1 co-fine-tuning (author-reported)