CHAI – Compliant Human-centered Adaptive Interaction through Diffusion-based Language Trajectory Transformer

Junhui Huang, Xingguang Duan, A. Bucker, Changsheng Li, L. Figueredo

TL;DR: CHAI (Compliant Human-centered Adaptive Interaction), a novel language-driven framework for real-time modulation of a robot’s kinematics and mechanical compliance in real-world environments, introduces on-the-fly language-driven impedance (compliance) modulation along both translational and rotational directions.

We present CHAI (Compliant Human-centered Adaptive Interaction), a novel language-driven framework for real-time modulation of a robot’s kinematics and mechanical compliance in real-world environments. By interpreting natural language instructions and visual context through pre-trained vision-language encoders, CHAI combines a transformer-based geometry encoder with a conditional diffusion model to iteratively refine a nominal kinematic trajectory and its associated compliance profile. CHAI introduces on-the-fly language-driven impedance (compliance) modulation along both translational and rotational directions—including motion-aligned and radial axes—executed through a passivity-aware and therefore stable Cartesian impedance controller. This capability is key for supporting compliant interaction, improving adaptability and reducing hazardous contact in physical interaction tasks. Comprehensive experiments demonstrate significant gains in scalability, adaptability, trajectory accuracy, and interactive behaviour over prior methods. Ablation studies further validate the contributions of stiffness control and multi-modal conditioning.