Data flywheels promise continuous improvement via insights from deployment data. Understanding how the deployed model succeeds and fails can steer subsequent data collection towards patching failures and reinforcing successes. However, robot deployments create large volumes of high-dimensional rollout data, which is challenging to understand and confidently ground data collection decisions in. To make deployment data easier to analyze and act on, we propose behavior graphs, which compress policy rollouts into a single interpretable, directed graph. We construct the graphs by learning to identify similar behavior modes across rollouts, connecting the behaviors temporally, and finally linking them to outcomes. In a user study on three real-world tasks, where participants analyzed rollouts to choose a data collection strategy, 16 of 18 preferred behavior graphs to rollout videos, reaching comparable conclusions with 40.3% lower perceived temporal demand and 37.6% lower effort. Beyond interpretability for deployment data, behavior graphs are capable progress predictors: across 43 datasets, they improve the rank correlation between predicted progress and true temporal order by 22.9% and reduce mean absolute error by 53.5% relative to the strongest baseline. As the number of deployed robots grows, behavior graphs lend structure to high-dimensional policy rollouts and make deployment data easier to interpret and reason about.
Figures & tables
Figure 1: Behavior graph construction. Left: Each trajectory’s state-action pairs are embedded, mean-pooled over temporal windows, and assigned to behavior modes by dimensionality reduction and clustering fit on all trajectories. Consecutive windows with the same mode are collapsed into segments, and a VLM captions each mode. Right: The resulting behavior sequences from all trajectories are merged into a single directed graph from a start node to success and failure outcomes, with edges weighted by empirical transition probabilities.
Figure 2: Behavior graph for the cooking task of our user study ( Sec. IV-A ). The larger the node, the higher the visitation count. Node colors indicate success likelihood (increasing from red to green), while all terminal nodes are annotated with their likelihoods. Branches with less than 2% likelihood are pruned to simplify visualization.
Figure 3: Tasks for the user study, in which participants analyze policy rollouts to choose which demonstrations to collect next. Keyframes of a successful rollout from left to right. (a) Set table : places a pumpkin, then a milk carton from the shelf; the training data omits the carton’s leftmost position. (b) Prepare tray : slides out a tray, places a bowl from the top or bottom shelf onto it, and grasps a fork, requiring precise grasps. (c) Cooking : a long-horizon task that clears a ladle, adds a corn, carrot, and chili to the pot, and closes the lid.
Figure 4: Targeted data helps only when it addresses the policy’s most frequent failure mode. Estimated success rates [ 27 ] of the baseline policy and of policies retrained with 30 additional demonstrations for each candidate data-collection strategy, on three real-world DROID tasks. In Set table , only demonstrations with the milk carton on the left (the position missing from the baseline training data) improve performance. In Prepare tray , bottom-shelf bowl demonstrations improve over the baseline. In Cooking , no single-subtask strategy makes a significant difference. Violins show the Bayesian posterior over each policy’s success rate (uniform Beta prior). Dots and bars mark the posterior mean and 95% credible interval. Letters above each column [ 28 ] summarize pairwise significance tests within a task [ 29 ] : two policies that share a letter are not significantly different; policies with no letter in common are. Tests follow [ 29 ] with family-wise error rate controlled at 5%, and 60 policy rollouts to estimate each posterior. Below, histograms show the strategies participants chose with the graph and with the video browser, which are similar.
Figure 5: Behavior graphs make rollout analysis less demanding without changing the decisions users reach. Responses from 18 participants. (a) When free to choose a display for the third task, 16 of 18 participants picked the behavior graph over the video browser. (b-d) Self-reported ratings (0-100, in steps of 20) after analyzing a task with each display: the graph lowers perceived temporal demand (b) and effort (c), while confidence in the chosen data-collection strategy is unchanged (d). Boxes span the interquartile range with the median marked, whiskers span min/max values; dots are individual ratings. (e) The success rate regret (i.e., the gap between the best retrained policy in Fig. 4 and the one the participant selected) is similar for graph and videos. Arrows indicate whether higher ( ↑ ) or lower ( ↓ ) is better. Tab. I reports the corresponding effects adjusted for task, stage, and experience.
Outcome
Graph
Videos
Difference [95% CrI]
Temporal demand ↓
38.7
64.8
−26.1[−38.4,−14.0]
Effort ↓
38.1
61.0
−22.9[−37.7,−7.8]
Confidence ↑
61.3
58.8
+2.5[−9.8,+15.5]
Success rate regret ↓
7.9
7.0
+0.9[−6.1,+8.5]
Table I : Effect of the behavior graph after adjusting for task, study stage, participant experience, and individual differences in scale use. We report estimated marginal means per visualization from the mixed-effects model and their difference. The reductions in perceived temporal demand and effort persist under adjustment, confidence and success rate regret are unchanged.
Domain (#Datasets)
Method
VOC ↑
Kendall ↑
MAE ↓
AUC ↑
LIBERO (6)
Graph (ours)
0.912
0.864
0.082
0.807
MLP
0.940
0.861
0.068
0.864
Robometer-4b
0.799
0.650
0.154
0.860
TOPReward
0.433
0.329
0.259
0.664
Real-world (19)
Graph (ours)
0.948
0.930
0.047
0.904
MLP
0.731
0.564
0.173
0.782
Table II : Progress-estimation metrics over 43 datasets spanning 4 domains. AUC is undefined for the two Robomimic datasets in which every rollout succeeds (Can-PH, Lift-PH), so the Robomimic and Average AUC entries are over 8 and 41 datasets. On average, graph-based progress estimation outperforms all baselines in the fidelity of the prediction. In AUC, which measures success classification induced by the progress estimates, Robometer outperforms on average.
Representation
Reduction
F1(0.75)↑
IoU ↑
Proprioception
none
0.093 ± 0.075
0.523 ± 0.061
PCA
0.120 ± 0.121
0.547 ± 0.098
UMAP
0.182 ± 0.022
0.598 ± 0.033
mean
0.132 ± 0.072
0.556 ± 0.063
Policy embeddings
none
0.264 ± 0.108
0.633 ± 0.048
PCA
0.254 ± 0.121
0.644 ± 0.059
Table III: Ablation study results for subtask segmentation. We report mean ± std over 3 datasets. Each mean row averages every dataset’s score over the three reductions, then reports the mean ± std of those averages across datasets. We report F1 at 0.75 IoU threshold and the mean IoU over matched segments. On average, policy embeddings outperform proprioception and frozen DINOv3 features. While there is no clear quantitative conclusion for reductions, UMAP performs more consistently than the others and is what we use for our other experiments.
Autonomous mobile robots require spatial representations that capture geometry and semantics while staying sparse enough for planning in large, unknown environments. Graphs provide a natural representation for this purpose and underpin both classical planning methods and a growing class of reinforcement-learning (RL) policies. However, RL training needs hundreds to thousands of parallel environments, each constructing and updating its own graph. Existing graph libraries are typically designed for either single-robot deployment or fixed graphs, and CPU-based graph construction limits scalable training to only a few dozen environments. We introduce GGraph, an open-source GPU-accelerated graph library for both robot deployment and large-scale learning. GGraph supports real robots through ROS 2 interfaces and vectorized GPU-based policy training, with built-in support for tasks including 3-D navigation, exploration, and loop-closure integration. For single-robot operation, GGraph updates graphs in milliseconds. At scale, it enables graph-based exploration policies to train across hundreds of parallel robots, reducing wall-clock training time by one order of magnitude. Across simulation and real-world experiments, these results establish GGraph as a unified graph infrastructure for both large-scale robot learning and deployment.
Recent advances in learning-based robot manipulation have produced policies with remarkable capabilities. Yet, reliability at deployment remains a fundamental barrier to real-world use, where distribution shift, compounding errors, and complex task dependencies collectively undermine system performance. This dissertation investigates how the reliability of learned robot policies can be improved at deployment time through mechanisms that operate around them. We develop three complementary classes of deployment-time mechanisms. First, we introduce runtime monitoring methods that detect impending failures by identifying inconsistencies in closed-loop policy behavior and deviations in task progress, without requiring failure data or task-specific supervision. Second, we propose a data-centric framework for policy interpretability that traces deployment-time successes and failures to influential training demonstrations using influence functions, enabling principled diagnosis and dataset curation. Third, we address reliable long-horizon task execution by formulating policy coordination as the problem of estimating and maximizing the success probability of behavior sequences, and we extend this formulation to open-ended, language-specified tasks through feasibility-aware task planning. By centering on core challenges of deployment, these contributions advance practical foundations for the reliable, real-world use of learned robot policies. Continued progress on these foundations will be essential for enabling trustworthy and scalable robot autonomy in the future.
Christopher Agia
THE DEPARTMENT OF COMPUTER SCIENCE AND THE COMMITTEE ON GRADUATE STUDIES OF STANFORD UNIVERSITY
Existing robot policies based on learned visual embeddings lack explicit structure and are sensitive to visual distractions. Thus, the representations that drive their behaviour are often opaque, making their decision-making process difficult to interpret. To address this, we introduce Structured Image Representations (SIR), a method that leverages Scene Graphs (SGs) as an intermediate representation for robot policy learning. Our approach first constructs a fully connected graph, using image-derived features as initial node representations. Then, a module learns to sparsify this graph end-to-end, creating a task-relevant sub-graph that is passed to the action generation model. This process makes our model intrinsically explainable. Evaluations on RoboCasa show that our sparse graph policies outperform image-based baselines on average with 19.5% vs 14.81% success rate. Most importantly, we show that the learned sparse graphs are a powerful tool for model analysis. By analysing when the model's sub-graph deviates from human expectation, such as by including distractor nodes or omitting key objects, we successfully uncover dataset biases, including spurious correlations and positional biases. https://github.com/intuitive-robots/SIR_Model
Paul Mattes, Jan Schwab, Jens Bosch +5
Intuitive Robots Lab, Karlsruhe Institute of Technology, Germany · Robotics Institute Germany