Robots that can adapt like animals (1407.3501v4)

Abstract: As robots leave the controlled environments of factories to autonomously function in more complex, natural environments, they will have to respond to the inevitable fact that they will become damaged. However, while animals can quickly adapt to a wide variety of injuries, current robots cannot "think outside the box" to find a compensatory behavior when damaged: they are limited to their pre-specified self-sensing abilities, can diagnose only anticipated failure modes, and require a pre-programmed contingency plan for every type of potential damage, an impracticality for complex robots. Here we introduce an intelligent trial and error algorithm that allows robots to adapt to damage in less than two minutes, without requiring self-diagnosis or pre-specified contingency plans. Before deployment, a robot exploits a novel algorithm to create a detailed map of the space of high-performing behaviors: This map represents the robot's intuitions about what behaviors it can perform and their value. If the robot is damaged, it uses these intuitions to guide a trial-and-error learning algorithm that conducts intelligent experiments to rapidly discover a compensatory behavior that works in spite of the damage. Experiments reveal successful adaptations for a legged robot injured in five different ways, including damaged, broken, and missing legs, and for a robotic arm with joints broken in 14 different ways. This new technique will enable more robust, effective, autonomous robots, and suggests principles that animals may use to adapt to injury.

• The paper introduces the Intelligent Trial and Error (IT&E) algorithm that uses a precomputed behavior-performance map and Bayesian optimization to quickly recover from damage.
• The methodology was validated on a hexapod and a robotic arm, showing significant speed improvements and effective compensation across diverse damage conditions.
• The approach mimics biological adaptation, offering practical insights for enhancing robotic resilience in dynamic and unpredictable environments.

Intelligent Trial and Error for Rapid Robot Damage Recovery

The paper "Robots that can adapt like animals" by Antoine Cully, Jeff Clune, Danesh Tarapore, and Jean-Baptiste Mouret presents an innovative approach to enhancing robot robustness through an intelligent trial and error algorithm. The proposed method allows autonomous robots to adapt quickly to damage without requiring detailed self-diagnosis or pre-specified contingency plans.

Intelligent Trial and Error Algorithm

The core contribution of the paper is the Intelligent Trial and Error (IT&E) algorithm, which comprises two major steps:

1. Behavior-Performance Map Creation: Before deployment, a robot uses a novel algorithm called MAP-Elites to create a detailed behavior-performance map. This map is generated in simulation and contains high-performing behaviors across a range of conditions. It represents the robot's prior knowledge about viable behaviors.
2. Adaptation Phase: When the robot sustains damage, it uses its behavior-performance map to guide a trial-and-error learning process. This phase employs a Bayesian optimization approach, specifically the map-based Bayesian Optimization Algorithm (M-BOA), to select and test promising behaviors swiftly. The robot updates its predictions based on the outcomes of these experiments and converges on an effective compensatory behavior.

Experimental Validation

The authors validated their approach with two experimental setups: a hexapod robot and an 8-joint robotic arm. For the hexapod robot, the algorithm was tested under six conditions, including undamaged and multiple damage scenarios (e.g., missing legs). The robotic arm underwent 14 different damage conditions, demonstrating IT&E's versatility.

Hexapod Experiment

Key results from the hexapod experiments include:

• In the undamaged condition, IT&E generated gaits that were 30% faster than the classic hand-designed tripod gait.
• For damaged conditions, IT&E quickly discovered compensatory behaviors that restored a significant portion of the robot's functionality, achieving speeds multiple times faster than the reference gait for each damage type.
• The adaptation process generally required less than 30 seconds for undamaged or mildly damaged conditions and approximately one minute for more severe damage.

Robotic Arm Experiment

In the robotic arm experiments, IT&E demonstrated similar success:

• The algorithm identified compensatory behaviors in less than two minutes for all tested conditions.
• The adaptation process required fewer than 10 trials to effectively compensate for various joint damages.

Theoretical and Practical Implications

The theoretical implications of this research are profound. By integrating intelligent trial and error with pre-computed behavior-performance maps, the approach addresses the curse of dimensionality that plagues many traditional learning algorithms. The process mimics biological adaptation, where experiences guide the selection of compensatory behaviors without exhaustive trials.

Practically, this work promises significant advancements in the field of autonomous robotics. Robots equipped with IT&E can operate reliably in dynamic and unpredictable environments, enhancing their utility in sectors such as search and rescue, disaster response, and space exploration.

Future Directions

Future developments in this area may involve:

• Generalization to Diverse Robots and Environments: Extending the approach to various robotic platforms and more complex environments.
• Enhanced Behavior Descriptors: Exploring alternative and richer behavior descriptors to capture more nuanced robot-environment interactions.
• Algorithmic Efficiency: Further optimization of the IT&E algorithm to reduce computing time and resource usage.

In conclusion, the Intelligent Trial and Error algorithm represents a significant step forward in making autonomous robots more adaptive and resilient to damage, drawing clearly from principles observed in natural organisms. The comprehensive experiments and strong numerical results presented in the paper underline the algorithm's efficacy and potential for broad application in real-world scenarios.