Learning Sidewalk Autopilot from Multi-Scale Imitation with Corrective Behavior Expansion
A new imitation learning framework teaches robots to recover from errors using corrective behavior expansion.
A research team led by Honglin He and Yukai Ma has published a paper titled 'Learning Sidewalk Autopilot from Multi-Scale Imitation with Corrective Behavior Expansion' (arXiv:2603.22527), presenting a breakthrough framework for teaching robots to navigate complex sidewalks. The work tackles a critical problem in last-mile micromobility: current imitation learning (IL) methods, which train AI by watching human demonstrations, suffer from compounding errors and poor generalization when faced with real-world unpredictability. Their solution attacks the issue from two angles—data and model architecture—to create a more robust autonomous navigation policy.
On the data side, the team's key innovation is 'corrective behavior expansion.' Instead of relying solely on standard teleoperation datasets of perfect driving, they intentionally augment the data with diverse recovery actions and sensor variations. This teaches the AI policy how to recognize and correct its own mistakes, much like a human learner practices recovering from a skid. On the model side, they introduce a multi-scale IL architecture. This system uses horizon-based trajectory clustering and hierarchical supervision to simultaneously capture short-horizon interactive behaviors (like avoiding a pedestrian) and long-horizon goal-directed intentions (like navigating to a specific block).
The result, validated through real-world experiments, is a significant leap in robustness and generalization for sidewalk navigation AI. The framework moves beyond simply mimicking pre-recorded paths and enables adaptive, fault-tolerant behavior in diverse urban scenarios. This research, shared on the arXiv preprint server, represents a meaningful step toward deploying reliable autonomous delivery robots, scooters, or assistive devices in the chaotic, pedestrian-filled environments where they are needed most.
- Proposes 'corrective behavior expansion' to augment training data with recovery actions, teaching the AI to fix its own errors.
- Uses a multi-scale architecture to learn both short-term interactions and long-term navigation goals simultaneously.
- Demonstrates significantly improved robustness and generalization in real-world sidewalk tests compared to standard imitation learning.
Why It Matters
Enables more reliable autonomous last-mile delivery and personal mobility robots in complex, unpredictable urban environments.