Bi-HIL: Bilateral Control-Based Multimodal Hierarchical Imitation Learning via Subtask-Level Progress Rate and Keyframe Memory for Long-Horizon Contact-Rich Robotic Manipulation

A team of researchers has introduced Bi-HIL, a novel AI framework designed to solve one of robotics' toughest challenges: long-horizon, contact-rich manipulation. Tasks like assembling components or handling flexible materials require both high-level planning for sequences of actions and low-level, force-sensitive control for physical interaction. Traditional AI policies often fail here, struggling to coordinate over long timeframes or react to unstable contact forces. Bi-HIL addresses this with a hierarchical architecture that explicitly models the progression within each subtask, using a 'subtask-level progress rate' to guide the robot's phase of operation.

The core innovation lies in combining this hierarchical planning with bilateral control-based imitation learning. Bilateral control allows the robot to learn from human demonstrations that include force and haptic feedback, making it 'force-aware.' Bi-HIL stabilizes this learning by integrating a 'keyframe memory' that stores crucial states from successful demonstrations. This memory, conditioned by the progress rate, helps both the high-level task planner and the low-level controller stay synchronized and recover from errors. In evaluations on real robotic hardware for both single-arm and dual-arm tasks, Bi-HIL demonstrated more robust and consistent performance compared to non-hierarchical (flat) policies and ablated versions of its own system.

The results underscore a significant technical insight: for robots to reliably perform complex, real-world chores, AI must explicitly model the temporal structure of subtasks while being finely attuned to physical forces. This moves beyond treating a task as one long, monolithic action and instead breaks it into manageable phases with clear progression cues. The framework, detailed in the arXiv preprint 2603.13315, represents a step toward robots that can autonomously handle the nuanced, multi-step physical work common in manufacturing, logistics, and home environments.