Business Context

RoboCourier is training an imitation-learning policy for low-speed warehouse cart navigation across 12 fulfillment centers. A behavior cloning model trained only on expert demonstrations performs well offline but degrades sharply during live rollouts because small mistakes push the cart into states the expert dataset rarely contains.

Dataset

You are given logged state-action trajectories from a human teleoperator and additional states collected from policy rollouts that can be re-labeled by the expert. The goal is to implement and evaluate a DAgger-style training loop that reduces compounding errors versus standard behavior cloning.

• Size: 420K state-action pairs from expert trajectories, plus up to 180K policy-visited states available for iterative expert relabeling
• Target: Predict expert action at each state; either discretize into 9 action classes or model two continuous controls
• Class balance: Highly skewed toward “go straight / low steering correction”; recovery maneuvers are rare (<6%)
• Missing data: 3% intermittent sensor dropouts in lidar/depth features; some route-context features missing for newly mapped aisles

Success Criteria

A strong solution should show that DAgger improves closed-loop performance over behavior cloning by reducing intervention rate and cumulative trajectory deviation. Good enough means at least a 25% reduction in interventions per 1000 meters while keeping inference under 20 ms per step.

Constraints

• Expert relabeling is expensive, so each DAgger round can add at most 30K new states
• The operations team wants a model simple enough to debug and retrain weekly
• Evaluation must distinguish offline imitation loss from online rollout quality

Deliverables

1. Train a behavior cloning baseline and a DAgger-enhanced policy
2. Define the aggregation strategy for newly visited states and expert labels
3. Propose features, preprocessing, and a leakage-safe train/validation/test split by site or episode
4. Evaluate both offline and closed-loop metrics, and explain why DAgger addresses compounding error
5. Provide production-ready training code and thresholding/monitoring recommendations for deployment