Grant

🤝 Effective human-robot collaboration often requires robots to infer human intent, coordinate their timing, and know not only 𝘸𝘩𝘢𝘵 action to take, but 𝘸𝘩𝘦𝘯 to assist. In this work, we study whether modern Vision-Language-Action models can learn collaborative behavior directly from demonstrations, without hand-engineered intent recognition or planning pipelines. Introducing “𝙇𝙚𝙖𝙧𝙣𝙞𝙣𝙜 𝙩𝙤 𝘼𝙨𝙨𝙞𝙨𝙩: 𝘾𝙤𝙡𝙡𝙖𝙗𝙤𝙧𝙖𝙩𝙞𝙫𝙚 𝙑𝙇𝘼𝙨 𝙛𝙤𝙧 𝙄𝙢𝙥𝙡𝙞𝙘𝙞𝙩 𝙃𝙪𝙢𝙖𝙣-𝙍𝙤𝙗𝙤𝙩 𝘾𝙤𝙡𝙡𝙖𝙗𝙤𝙧𝙖𝙩𝙞𝙤𝙣,” with Letian Li, Alex Cuellar, and @hagenowrobotics. ✨What we found: ✅ VLAs can learn collaborative behavior across real-world tasks involving sequential collaboration, handovers, and implicit human intent. 🚨 But collaboration exposes a timing failure mode we call 𝙙𝙚𝙢𝙤𝙣𝙨𝙩𝙧𝙖𝙩𝙞𝙤𝙣 𝙖𝙘𝙩𝙞𝙤𝙣 𝙡𝙚𝙖𝙠𝙖𝙜𝙚. Because many VLAs predict chunks of future actions, a training chunk can cross from waiting into handover. As a result, the robot may begin assisting before the human is ready, causing premature actions, drift, and failures. To address this, we introduce an inference-time steering method that keeps the robot near demonstrated waiting states until the human shows readiness, requiring no retraining. Across our experiments, steering reduced premature handovers by 96.7% in our false-commitment study. In a 16-participant user study, steering enabled smoother, longer-horizon robot behavior while reducing task time and avoiding premature assistance. For more details on the approach, experiments, and results: 📄 Paper: arxiv.org/abs/2606.12475 🎥 Video: youtu.be/HRycLQUMZok We are grateful to @tutorintel for donating the robot hardware used in this project.