Embodiment-Aware Intent Transfer for Hazardous Labour Substitution: From Simulation to Indian Field Conditions

Abstract

Transfer learning for robotic manipulation typically assumes a consistent sim-to-real gap characterized by visual and dynamics differences between simulator and target environment. We identify a distinct and underexplored transfer challenge in hazardous labour substitution: embodiment-aware intent transfer, where the source demonstrations come from human workers performing tasks with their bodies, and the target is a robotic system with fundamentally different morphology. In the context of Indian sewer maintenance, workers use their hands, feet, and improvised tools to navigate, extract, and deposit material in irregular confined spaces. We propose a framework that decomposes human task demonstrations into embodiment-invariant intent representations, which are then re-grounded in the robot's action space through a learned mapping. We evaluate this approach using data from our ethical wearable data collection programme with Indian sanitation workers and demonstrate a 34% improvement in task completion over direct behavioral cloning from human demonstrations. The paper discusses the broader implications of intent transfer for deploying robots in labor contexts where the incumbent workers, not engineers, are the domain experts.

Keywords