- Erik Bauer and colleagues posted a paper to arXiv on October 30, 2026, introducing Behaviour Graphs, an unsupervised method for interpreting robot deployment data that requires no manual labelling of robot behaviour.
- In a user study on three real-world tasks, participants reached comparable data collection decisions with 40.3% lower perceived temporal demand and 37.6% lower effort than with rollout videos, and 16 of 18 preferred the graphs.
Erik Bauer and eight co-authors posted a paper to arXiv on October 30, 2026 proposing behaviour graphs, which compress robot policy rollouts into a single directed graph of behaviour modes, their transitions and their outcomes. The approach is unsupervised, meaning it requires no manual labelling of robot behaviour to work.
From Rollouts to Policy Embeddings
Robot deployments produce large volumes of high-dimensional rollout data, which the paper says is hard to understand and hard to base data collection decisions on. Behaviour graphs are meant to address that by showing how a deployed policy succeeds and fails, so teams can patch failures and reinforce successes. To build one, the authors embed each trajectory’s state-action pairs and average those embeddings over temporal windows, producing one policy embedding per rollout segment. Those segment embeddings are then reduced and clustered across all trajectories to find the behaviour modes described next.
Behaviour Modes as Nodes
Once each rollout is split into segments and embedded, the framework has to turn those vectors into something a person can read. According to the paper, it does this by clustering the segment embeddings from all rollouts together, so that similar behaviours in different rollouts land in the same cluster. Each cluster becomes a node in the graph and stands for a shared behaviour mode, which the authors describe as semantic. Because the clustering is fit on every trajectory at once, a single graph can summarize a whole set of rollouts instead of describing each run separately. The construction is unsupervised, so no human-annotated examples define the modes. The paper’s Figure 2 shows the resulting graph for the cooking task in its user study.
Building the Graph
Directed edges connect each behaviour mode to its successors and to rollout outcomes, weighted by the empirical transition probabilities observed across rollouts. A behaviour that frequently leads to a failure outcome would therefore carry a heavily weighted edge into that outcome, which is the kind of pattern the graph is meant to expose. Each rollout’s sequence of behaviours starts at a shared start node and ends with an edge into that rollout’s outcome, so the graph captures not just what a robot typically does but where rollouts end.
User Study and Progress Prediction
Participants analysed rollouts to choose a data collection strategy using either behaviour graphs or rollout videos. Sixteen of 18 preferred the graphs, and they reached comparable conclusions with 40.3% lower perceived temporal demand and 37.6% lower effort.
The paper also tests the framework as a progress-estimation tool. Across 43 datasets, the paper reports that behaviour graphs improved the rank correlation between predicted progress and true temporal order by 22.9% and reduced mean absolute error by 53.5% relative to the strongest baseline. The authors say behaviour graphs could help as the number of deployed robots grows. The user study was small, with 18 participants, and the demand and effort figures are self-reported.
