Paper review · October 24, 2026
Learning robot action priors from synthetic, non-robot data
Learning Robot Behavior Priors from Non-Robotic Synthetic Data
By pretraining a Transformer only on mathematical control abstractions such as MDPs, contextual bandits and linear dynamical systems, the authors obtained in-context performance well above random on real robot action prediction without ever showing the model robot data.
Problem
Robot learning depends heavily on expensive real robot data, and simulators merely shift that burden onto domain-specific engineering. The authors ask whether prior structure useful for robot action prediction can be learned from cheap procedural synthetic data with no embodiment at all.
Method
Following the Prior-Data Fitted Network (PFN) approach, they train a Transformer that, given a state-action context, predicts held-out actions in a single forward pass. Continuous actions are discretized into 64 bins with K-means. There are three synthetic data generators: MDPs with rewards and transitions sampled from GPs and solved by value iteration; contextual bandits covering single-step state-action choices; and linear dynamical systems (LDS) with strong temporal autocorrelation. Feature order is randomly shuffled each time. This is compared with TabPFN's general-purpose tabular prior and with a baseline trained on real robot data using the same compute.
Results
- On a VMAS simulated multi-agent task (9 actions), the control prior model reached 0.58 accuracy, higher than the tabular prior (0.49).
- Average accuracy across four real robot datasets (A2D2, DROID, flight, GrandTour) is 0.166 for the control prior and 0.039 for the tabular prior (random is 1/64≈0.016). RCD improved, 0.706 vs. 0.905.
- In ablations, LDS was the strongest single prior and performed similarly to the full mixture on the real robot set. The authors interpret this as local temporal continuity transferring rather than long-term value structure.
- The authors report that baselines trained on real robot data were much stronger on the datasets used for training, but the proposed model came out ahead on the held-out set of tasks.
Limitations
Absolute accuracy over 64 classes is still low, at around 0.12–0.24, and the authors themselves describe the approach as a complement to, not a replacement for, training on real data. The evaluation covers only offline action prediction and does not measure actual robot control performance. The setup discretizes continuous actions and uses only state vectors, with no vision input.
Why it matters
Where collecting robot data is expensive, this offers evidence that cheap, mathematically generated synthetic data can be useful as a pretraining prior. It also gives clues about which kinds of synthetic structure (temporal continuity) transfer to real robots.
Key terms
- PFN(Prior-Data Fitted Network)
- A Transformer trained on synthetic datasets sampled from a prior distribution so that it can look at the context of a new dataset and make predictions immediately, without further training. TabPFN is a well-known example.
- In-context Learning
- The ability to answer a new query using only the examples (context) given in the input, without updating any weights.
- Linear dynamical system (LDS)
- A system in which the next state is a linear transformation of the current state plus an action-specific displacement and noise. It mimics states that change smoothly, as in physical systems.
- RCD(Relative Centroid Distance)
- A metric that divides the distance between the center of the predicted action bin and the center of the correct one by the expected distance under random guessing. 0 means perfect; 1 means random-level.
Original
Authors: Mateusz Odrowaz-Sypniewski, Michael Amir, Jasmine Bayrooti, Ajay Shankar, Amanda Prorok · Posted 2026-10-22 · License arXiv 비독점 배포 라이선스 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.