Recoleta
GitHub
Source note

Towards a Data Flywheel for Embodied Intelligence in Logistics

Authors
Anlan Yu, Zaishu Chen, Zhiqing Hong, Daqing Zhang
Source
arXiv
Published
Collection
Embodied AI
World ModelsRobot Data ScalingImitation LearningIndustrial LogisticsSynthetic Recovery Data
Open arXivMarkdown source

Summary

The paper proposes a logistics data flywheel that turns daily warehouse operations, robot logs, demonstrations, and failure cases into training data for embodied policies. Its concrete result is WM-DAgger, which uses an action-conditioned World Model to generate recovery trajectories for imitation learning.

Problem

  • Logistics robots face long-tail parcel manipulation cases under throughput and reliability constraints, and lab demonstrations do not cover enough failures or recovery behaviors.
  • Human-in-the-loop DAgger-style recovery labeling is costly at industrial scale.
  • Existing operational data is heterogeneous, so it must be aligned into observation-action-outcome sequences before it can train World Models or policies.

Approach

  • Train an action-conditioned World Model from limited expert demonstrations plus about 5 minutes of play data.
  • Generate candidate video-action recovery trajectories by predicting future eye-in-hand observations for robot actions.
  • Use Corrective Action Synthesis to perturb expert trajectories toward nearby out-of-distribution states and guide actions back toward expert behavior.
  • Use Consistency-Guided Filtering to reject generated rollouts whose terminal frames disagree with matched real demonstration frames.
  • Aggregate filtered synthetic recovery data with real demonstrations to train imitation policies; future work adds unlabeled operational videos, automation logs, and deployment feedback.

Results

  • Soft Bag Pushing (5-shot): 5 real demonstrations plus 1,500 generated trajectories reached 93.3% success, versus 26.7% for standard BC and 40.0% for DMD.
  • Soft Bag Pushing (20-shot): 96.7% with WM-DAgger, versus 30.0% BC and 56.7% DMD, using 20 real demonstrations and 1,500 generated trajectories.
  • Pick-and-Place (Seen): 84.4% with WM-DAgger, versus 11.1% BC and 32.2% DMD.
  • Pick-and-Place (Unseen): 70.0% with WM-DAgger, versus 5.0% BC and 11.8% DMD.
  • Ballot Insertion and Towel Folding reached 73.3% and 46.7% success with WM-DAgger, compared with 13.3% and 0.0% for BC.
  • The broader logistics flywheel beyond WM-DAgger is proposed work; the excerpt reports no production-scale deployment metric.

Link