Key Concepts
- Fluid Physical Interaction: Seamless adaptation, exchange of power, energy, and information between humans and robots.
- Dynamical Systems Representation: Modeling robot motion as continuous functions allowing for compliant and adaptive behavior.
- Feedback at All Levels: Integrating feedback loops into high-level decision making, motion planning, and low-level control.
- Human-Aware Control: Incorporating human limitations and capabilities into robot control algorithms for safe and intuitive interaction.
- Dynamical System-Based Obstacle Avoidance: Utilizing the inherent flow properties of dynamical systems to navigate around obstacles.
- Viability Preserving Control: Ensuring robot safety by preventing entry into unsafe states (joint limits, collisions) while maintaining passivity.
Achieving Fluid Physical Interaction: Foundations (Part 1)
The research focuses on achieving “fluid physical interaction” – a seamless adaptation, exchange of power, energy, and information between humans and robots. This necessitates robots that are simultaneously stable, safe, flexible, adaptive, predictable, and compliant. Traditional robot control pipelines, with their limited feedback, are insufficient. The core methodology involves tightly coupled control, estimation, and learning algorithms with feedback integrated at all levels – high-level decision making, motion planning, and low-level control.
A key principle is representing motion plans as dynamical systems (functions f(x)) converging to attractors, rather than tracking pre-planned trajectories. This allows for more natural, compliant movement. An elastic motion policy combines dynamical systems with task constraints (SC3 frames) and utilizes a particle filter to estimate the parameters of the dynamical system online, adapting to changes in the environment or human intent. The confidence level of this intent estimation modulates the robot’s impedance. Manipulability ellipsoids representing the human’s reachable workspace shape the noise of the particle filter, ensuring feasible human motion. Experiments demonstrated successful adaptation to changing task parameters (e.g., object location) with a single demonstration, estimating human intent at 20 Hz.
Advancing Compliant Interaction & Safety (Part 2)
Building on the foundation, this segment details advancements in compliant and safe human-robot interaction. Human-awareness is crucial; robots must consider human limitations when estimating dynamical systems. Manipulability ellipsoids of the human skeleton are used to shape the noise within the particle filter, ensuring the robot’s estimated system is feasible for the human. Capability-aware control with low-level torque controllers addresses joint limits and workspace constraints.
Experiments involved a human attempting to reach SC3 poses on a screen, with the robot providing assistive force. The proposed method required less effort from the human than classical admittance control and performed comparably to goal-aware control, despite lacking prior goal information. This demonstrates online, real-time intent inference.
Obstacle avoidance is framed as a natural extension of the dynamical system representation, leveraging the concept of flow. Obstacles reshape these flows using a modulation matrix derived from a continuously differentiable representation of the obstacle boundary (gamma(x)). This shrinks velocities towards obstacles and redirects them tangentially, ensuring smooth avoidance. This approach, presented at TRO, handles complex geometries and generates exit strategies for concave obstacles.
The modulation approach is mathematically linked to Control Barrier Functions (CBFs), but addresses their local minima issues with modulated control barrier functions incorporating exit strategies and utilizing Model Predictive Path Integral (MPPI) for trajectory prediction.
Finally, viability preserving passive torque control addresses maintaining passivity while preventing unsafe states. A function (gamma(q, q_dot)) learns viable states based on offline simulations and is used in a simple quadratic program (QP) to compute safe torques, avoiding complex calculations required by higher-order CBFs. A demonstration showcased smooth adaptation to external perturbations and collision avoidance. Future work includes estimating dynamic mismatch parameters and combining this with flow matching policies, with a current project focused on a robotic exercise system adapting to user muscle activity (measured via ultrasound).
Conclusion
This research presents a comprehensive approach to achieving fluid physical interaction between humans and robots. By representing motion as dynamical systems, integrating feedback at all control levels, and prioritizing human-awareness and safety through techniques like manipulability ellipsoid integration, dynamical system-based obstacle avoidance, and viability preserving control, the work demonstrates significant progress towards robots capable of seamless, adaptive, and safe collaboration with humans in complex, dynamic environments. The emphasis on learning from limited data and leveraging inherent system properties offers a promising path towards more robust and generalizable human-robot interaction systems.
AI summaries can miss context or contain errors. Check important details against the original video.
