Task-Oriented Robot-Human Handovers on Legged Manipulators

Why It Matters

Zero‑shot, task‑aware handovers lower human workload and accelerate deployment of mobile manipulators in dynamic workplaces, giving manufacturers a competitive edge.

Key Takeaways

• LLM reasoning enables zero-shot handover planning
• Texture transfer bridges visual gaps for object affordances
• Study shows 71% user preference over prior methods
• Reduces human regrasping effort in collaborative tasks
• Demonstrated on ETH Zürich legged manipulator platform

Pulse Analysis

Robot‑human handovers have long been a bottleneck in collaborative automation, often requiring extensive pre‑programming or costly data collection to teach robots how to pass objects safely. Traditional pipelines rely on geometric grasp planners that ignore the nuanced intent behind a handover, leading to awkward exchanges and increased human effort. Recent advances in large language models (LLMs) provide a pathway to embed semantic understanding directly into the decision‑making loop, allowing robots to infer the purpose of an object and select handover strategies that align with human expectations.

AFT‑Handover leverages this semantic power by using an LLM to generate affordance hypotheses for a target object, then refines those hypotheses with texture‑based affordance transfer that maps visual cues to functional grasps. This dual‑stage approach eliminates the need for task‑specific training data, achieving true zero‑shot performance. In a head‑to‑head user study, the framework outperformed the previous best method, with 71.43% of participants rating it higher due to smoother transfers and less need for post‑handover regrasping. The experimental validation on ETH Zürich’s legged manipulator demonstrates that the system can operate on mobile platforms, maintaining stability and precision even while the robot navigates uneven terrain.

The implications extend beyond academic labs. Industries such as logistics, construction, and healthcare increasingly deploy mobile manipulators that must interact fluidly with human workers. By reducing the programming overhead and improving handover ergonomics, AFT‑Handover can accelerate adoption of collaborative robots, lower training costs, and enhance workplace safety. Future research will likely explore scaling the LLM reasoning to multi‑object scenarios and integrating tactile feedback, further bridging the gap between semantic intent and physical execution in human‑robot teams.

Original Description

We present AFT-Handover, a framework that integrates large language model (LLM)-driven affordance reasoning with texture-based affordance transfer to achieve zero-shot, task-oriented robot-to-human handovers. In a comparative user study, our framework is preferred over the current state-of-the-art by 71.43% of the participants, reducing human regrasping effort and enhancing perceived task understanding. We demonstrate real-world task-oriented handovers on legged manipulators, highlighting the potential of integrating semantic reasoning with affordance transfer for robot-human handovers on mobile manipulators.

For more information, check out our paper (https://arxiv.org/abs/2602.05760) or come visit our talk at the International Conference on Human-Robot Interaction (HRI 2026)!

Authors: Andreea Tulbure, Carmen Scheidemann, Elias Steiner, Marco Hutter — Robotic Systems Lab, ETH Zürich