Research

The Soft Robotics Control Lab integrates the embodied intelligence of soft materials with artificial intelligence to make them move.

We seek to answer fundamental questions about balancing these two methods. When robots are designed to conform to their environments, how much must a control system create a precise motion? When soft materials are computationally difficult to model, how accurate do these models need to be for control? Although softer materials are intuitively more safe than rigid ones, how do we mathematically verify that safety? If softer robots are to assist in our everyday lives, how do we built them larger and stronger without sacrificing the benefits of softness?

We address these questions by four research directions: (1) Simplified Modeling, (2) Feedback Control and Safety, (3) Planning and Trajectory Generation, and (4) Design for Control and Interaction.

Videos on this page may take a while to load!

Simplified Modeling

Our group uses both physics-based methods and data-driven approximations to efficiently model complicated soft systems.

For smart thermal actuators, such as shape memory alloys (SMAs), even simple recurrent neural networks can learn hysteresis models from sensor data for predictions of dynamic motion. And if a soft robot is moving slowly, a helpful approximation is static equilibrium. We have used static beam bending approximations as part of control for soft robot manipulators, and optimization-based methods for flexible cable-driven robots.

Feedback Control and Safety

Robot safety is much simpler to express with artificial intelligence than embodied intelligence: if a control system can guarantee that a robot’s state remains within some bounds, we consider that to be safe. We have developed verifiably-safe control systems for thermally-actuated robots, keeping temperatures bounded even during human interactions.

These controllers have been implemented in a soft legged robot that can balance without overheating its actuators:

We have also shown that computational methods, such as robustness verification and model-predictive control, can balance imperfect models with improvements from feedback:

Planning and Trajectory Generation

For soft robots whose motions do not need to be exact, closed-loop feedback control may not be needed. Our lab has developed methods for calculation of open-loop trajectories for soft robot limbs that can mimic a human’s desired motion:

We have also developed planning algorithms for dynamic motions of aquatic soft robots, which can be used online:

Design for Control and Interaction

Implementing these control methods in hardware requires designs with sensing and actuation for the robot’s full state. Our recent work has developed soft walking robots with these capabilities, and our lab is working on scaling up designs for larger forces.