Graph-Reweighted Behavior Cloning with Human Interventions for Precise Robotic Manipulation
Abstract
Long-horizon precise manipulation produces mixed-quality data for post-training vision-language-action (VLA) models: even successful demonstrations and intervention-assisted rollouts may contain detours, hesitations, failed attempts, and imperfect recoveries. The key challenge is to identify which action chunks actually advance the task, rather than uniformly cloning or discarding imperfect trajectories. We propose Graph-Reweighted Behavior Cloning (GR-BC), an offline iterative post-training framework that builds a task-conditioned goal-reachability graph over demonstrations and iteratively collected rollouts, including both successful and failed episodes with human interventions. Reverse graph search computes minimal remaining steps to task success within the dataset support, and each action chunk is weighted by how much its execution reduces this potential. This consequence-based weighting reuses recoverable segments from failed trajectories, leverages interventions as recovery paths, and down-weights locally regressive chunks even inside successful trajectories. New rollouts are incorporated by appending them to the graph and rerunning search, without training a reward model, action-conditioned critic, or progress predictor. Across four precise manipulation tasks and three VLA backbones, GR-BC improves single-round supervised fine-tuning and reaches 91% success after a few offline iterations, without trajectory comparisons or progress labels. More details and videos are available at: https://gr-bc.github.io
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
The paper introduces a non-parametric, graph-based method for reweighting behavior cloning data in VLA models, effectively leveraging failed and intervened trajectories to improve precise robotic manipulation performance without requiring learned reward models or dense progress labels.
Methodology
The paper proposes Graph-Reweighted Behavior Cloning (GR-BC), a framework that addresses the problem of mixed-quality data in VLA post-training. The core innovation is the construction of a task-conditioned goal-reachability graph using multi-view MCR embeddings. By performing reverse Dijkstra search from successful terminal states, the method computes a "minimal remaining steps" potential for each state. Action chunks are then weighted based on how much their execution reduces this potential (consequence-based weighting). This approach is non-parametric regarding the progress signal (no learned reward model or critic), which is a significant advantage over methods requiring dense progress labels or VLM annotations. The method effectively handles failed trajectories by identifying recoverable segments that still lead toward success, rather than discarding them entirely. The integration of gripper-state consistency as a hard filter for graph edges is a practical and effective design choice to prevent physically incompatible state connections.
Results & evaluation
The experiments are conducted on four real-world precise manipulation tasks (Assembly Model, Graphics Card Assembly, Screw Tightening, Memory Stick Insertion) using a Flexiv Rizon 4s arm. The evaluation compares GR-BC against DAgger, AWR, and RA-BC (reward-aligned BC). GR-BC demonstrates superior success rates, reaching 91% on some tasks after a few offline iterations. The ablation studies on graph construction (temporal vs. similarity edges, gripper consistency) provide strong evidence for the necessity of the proposed graph structure. The progress estimation evaluation against baselines like SARM and ReWiND shows GRP's ability to detect local regressions without additional supervision. The inclusion of a backbone compatibility study (X-VLA, Pi0.5, SmolVLA) further validates the generality of the weighting mechanism.
Reproducibility
The paper provides detailed implementation details in the appendix, including hyperparameters, graph construction specifics, and evaluation protocols. The use of standard VLA backbones and clear definitions of the graph construction process enhances reproducibility. However, the reliance on specific visual encoders (MCR) and the complexity of the graph construction pipeline may pose challenges for replication without access to the specific codebase. The project page is provided, which likely contains code and videos.
Limitations
The method is limited by the quality of the dataset-supported reachability graph; incorrect visual neighbors can distort progress estimates. It assumes that shorter graph paths correspond to better behaviors, which may not always hold if force or collision risks are not accounted for in edge costs. Generalization to states outside the collected data support is limited. The computational cost of graph construction and search increases with dataset size, though this is noted to be less than neural retraining.
Broader impact
This work offers a practical solution for improving VLA models using real-world, imperfect data, which is a common scenario in robotics deployment. By avoiding the need for expensive human annotations or complex reward models, it lowers the barrier for effective post-training. The technique could be extended to other domains where sequential decision-making involves mixed-quality trajectories, such as autonomous driving or interactive AI agents. The paper introduces a non-parametric, graph-based method for reweighting behavior cloning data in VLA models, effectively leveraging failed and intervened trajectories to improve precise robotic manipulation performance without requiring learned reward models or dense progress labels.
Author of this paper? Send a correction.