RoboCat
Self-improving visual goal-conditioned decision transformer that masters new robot arms from 100–1000 demos.
About
RoboCat is DeepMind's 2023 answer to the data problem in robotics: a single goal-conditioned decision transformer that works across different robot arms and keeps improving itself. It trains on action-labelled visual experience from simulated and real arms with different observation and action spaces, then adapts to new tasks and embodiments with 100 to 1000 demonstrations. The neat trick is the self-improvement loop: a trained RoboCat generates its own successful trajectories, and those become training data for the next iteration. On tasks it had never seen, it started at about a 36% success rate and roughly doubled that through self-training, per the authors. It was evaluated on three real robot embodiments. What it is not: a downloadable product. DeepMind never released code or weights, so RoboCat lives on as a paper (arXiv 2306.11706, June 2023) and a widely cited design pattern.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper (arXiv 2306.11706): new-task success ~36% initially, roughly 2x after self-training; adapts with 100–1000 demos (author-reported)
- No independent replication or public benchmark numbers 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 2306.11706) with large-scale sim and real-robot evaluation across three embodiments; widely cited. Strengths: self-improvement loop demonstrated end to end, clear adaptation recipe with few demos. Docked for: (1) no public code or weights — nothing to inspect or reproduce; (2) every number is 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
- · Google DeepMind paper arXiv 2306.11706 (Jun 2023); visual goal-conditioned decision transformer
- · Self-improvement loop: trained model generates data for the next iteration; ~36% to ~2x on unseen tasks (author-reported)
- · No public code or weights — influential research design only