GR-BC: Graph Search Extracts Effective Actions from Imperfect Robot Demos

Problem: Learning from Imperfect Demonstrations

Precise robotic manipulation tasks like screw tightening require repeated micro-adjustments. Even skilled teleoperators rarely produce a single flawless trajectory; successful runs contain ineffective corrections, while failed attempts may include well-executed segments. This raises a core question for vision-language-action (VLA) post-training: what experience should the robot learn from when perfect demonstrations are unavailable?

Method: Graph-Reweighted Behavior Cloning (GR-BC)

GR-BC treats multiple demonstrations as a connected graph where effective action segments are identified through graph search and reweighting.

Step 1: Connecting Disparate Experiences

Nodes represent states encoded by a frozen MCR (manipulation-centric representation) visual encoder across three camera views plus gripper state. Two edge types link nodes:

Temporal edges connect adjacent nodes within the same trajectory at fixed sampling intervals; edge cost equals the number of frames crossed, so graph distance corresponds to physical time.

Similarity edges bridge independent trajectories when three conditions hold: mutual K-nearest neighbors in each camera view, consistent gripper state, and normalized visual distance below a threshold. Normalization uses the local scale of a single-step visual transition, making the threshold physically consistent across tasks and views.

Step 2: Reverse Search from Success

Adapting cost-to-go from reinforcement learning, GR-BC defines progress as the minimum graph-path cost from a state to any successful terminal state. After graph construction, reverse Dijkstra search computes the remaining-step count for each node. Results are linearly interpolated to every frame, aligning with dense VLA supervision.

Step 3: Dual-Check Weighting

Remaining steps are mapped to a 0–1 Graph-Retrieved Progress (GRP). Each action chunk must pass two checks: net progress (Δ) exceeds a threshold, and maximum backtrack (d) relative to the chunk start stays within a limit. Passing chunks receive a bounded positive weight; others get zero weight. Zero-weight segments contribute no gradient but remain in the graph for future iterations. The weight multiplies the standard VLA supervision loss without architectural changes.

Iterative Offline Loop

After each execution round, new successful, failed, and intervened trajectories are added to the data pool. The graph is expanded, progress recomputed, weights updated, and the next policy trained offline. New experiences can re-evaluate old segments without a separate reward model or progress predictor.

Experiments: Four Precise Manipulation Tasks

Real-robot tasks: four-hole part assembly, GPU insertion, dual-screw tightening, and DDR4 memory insertion (slot opening ~1 mm). Backbone: X-VLA, all methods starting from the same behavior cloning checkpoint.

GR-BC final success rates: 98%, 93%, 89%, 83% (average ~91%).

DAgger-style equal-weight aggregation baseline: average 64% (27 percentage points lower).

Additional GPU insertion experiments with SmolVLA and π₀.₅ show GR-BC outperforms their respective DAgger baselines, confirming backbone compatibility.

Progress Evaluation: Detecting Regressions within Success

Each task evaluated on 30 successful and 30 failed trajectories. Aggregated results: GRP success–failure terminal progress gap = 0.97; local deviation detection rate = 0.92. GRP not only distinguishes success from failure but also captures mid-trajectory regressions (local deviations, re-orientations, slips).

Conclusion and Outlook

GR-BC enables multiple imperfect attempts to collectively contribute effective experience, shifting data-value judgment from whole-trajectory success/failure to local action contribution. This provides a new paradigm for pre-training data cleaning, post-training data quality optimization, and progress-based data compliance assessment. For industrial millimeter-level insertion and assembly, the approach promises to help policies learn error correction and recovery, supporting more robust robotic manipulation.

Paper: https://arxiv.org/pdf/2610.20504 (CoRL 2026). Project page: https://gr-bc.github.io. Code: https://github.com/Gamma-Robot/GR-BC.