Bipedal Robot Capture Point Learning for Push Recovery
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing models for determining appropriate foot placement in bipedal robots are prone to significant errors, especially in robots with complex walking control systems and distributed mass, leading to inaccuracies in predicting stepping locations and balance recovery from pushes.
Innovation Solution
The development of techniques to determine and update 'capture points' on the ground surface that a biped robot can step to, using a combination of sensors, control modules, and learning methods to improve balance recovery, involving a stepping control module and stopping control module, with the use of a capture point memory and models like the Linear Inverted Pendulum model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If simple models are used to determine foot placement, then the control system is easy to implement, but the prediction accuracy of stepping location is poor
Solution Approach 1:
The system uses feedback from actual push events to update the capture point model. After each push recovery attempt, the robot compares the predicted capture point with the actual successful capture point (or the capture point that would have been successful), and uses this feedback to refine the model parameters, thereby improving prediction accuracy while keeping the model structure simple
Solution Approach 2:
The system changes the parameters of the capture point model based on learned experiences from push recovery events. By adjusting model parameters according to actual performance data, the system improves prediction accuracy without increasing the fundamental complexity of the model structure
2Measurement precision
If complex models are used to reduce errors, then the prediction accuracy improves, but the device complexity increases
Solution Approach 1:
Instead of using a complex model from the start, the system employs a simple initial model and continuously refines it through feedback from actual push recovery events. This iterative learning approach achieves high accuracy without requiring complex model structures
Solution Approach 2:
The capture point model serves itself by learning from its own performance in actual push recovery events. The system automatically updates its own parameters based on experience, eliminating the need for complex external calibration or sophisticated initial modeling
3Measurement precision
If accurate modeling parameters are used, then the prediction accuracy of stepping location improves, but small errors in parameters lead to significant errors in predicting desired stepping location
Solution Approach 1:
The system uses feedback from actual push recovery outcomes to continuously correct parameter errors. By comparing predicted versus actual capture points and adjusting parameters accordingly, the system becomes robust to initial parameter inaccuracies and maintains reliable predictions despite uncertainties in modeling parameters
Data Source
AI summary
A system and method is disclosed for controlling a robot having at least two legs, the robot subjected to an event such as a push that requires the robot to take a step to prevent a fall. In one embodiment, a current capture point is determined, where the current capture point indicates a location on a ground surface that is the current best estimate of a stepping location for avoiding a fall and for reaching a stopped state. The robot is controlled to take a step toward the current capture point. After taking the step, if the robot fails to reach a stopped state without taking any additional steps, an updated current capture point is determined based on the state of the robot after taking the step. The current capture points can be stored in a capture point memory and initialized based on a model of the robot.