Autonomous Excavator Trajectory Planning With Imitation-Led Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current trajectory planning methods for autonomous excavators face challenges in generating optimal and feasible trajectories, particularly in dynamic environments, due to limitations in model-based and learning-based approaches, such as lack of generalization, data efficiency, and difficulty in transferring simulation results to real-world applications.
Innovation Solution
A two-stage methodology integrating imitation learning (IL) and model-based trajectory planning, where expert trajectories are used to guide the generation of optimal trajectories that satisfy kinematic and hard constraints, using a stochastic trajectory optimization method (STOMP) to refine the initial trajectory and ensure feasibility and smoothness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If model-based trajectory planning is used, then trajectory feasibility is improved, but adaptability to different environments deteriorates
Solution Approach 1:
The patent introduces an imitation learning module as an intermediary between the environment and the model-based planner. This module learns from expert demonstrations and generates initial trajectories that are then refined by the model-based optimization, combining the adaptability of learning with the feasibility guarantees of model-based planning
Solution Approach 2:
The patent merges two different trajectory planning approaches: imitation learning (which provides adaptability) and model-based optimization (which provides feasibility guarantees). The combined approach uses the strengths of both methods to overcome their individual limitations
2Adaptability or versatility
If learning-based trajectory planning is used, then adaptability is improved, but data efficiency deteriorates
Solution Approach 1:
The patent performs preliminary action by pre-training the imitation learning module with expert demonstrations before deployment. This allows the system to learn from a limited set of expert trajectories and then generalize to new environments without requiring extensive additional data collection
3Productivity
If simulation training is used, then training efficiency is improved, but transfer to real-world deteriorates
Solution Approach 1:
The patent incorporates feedback from real-world expert demonstrations to refine the imitation learning model. By continuously incorporating real data and comparing with simulation results, the system reduces the simulation-to-reality gap while maintaining training efficiency
Data Source
AI summary
Presented herein are embodiments of a two-stage methodology that integrates data-driven imitation learning and model-based trajectory optimization to generate optimal trajectories for autonomous excavators. In one or more embodiments, a deep neural network using demonstration data to mimic the operation patterns of human experts under various terrain states, including their geometry shape and material type. A stochastic trajectory optimization methodology is used to improve the trajectory generated by the neural network to ensure kinematics feasibility, improve smoothness, satisfy hard constraints, and achieve desired excavation volumes. Embodiments were tested on a Franka robot arm equipped with a bucket end-effector. Embodiments were also evaluated on different material types, such as sand and rigid blocks. Experimental results showed that embodiments of the two-stage methodology that comprises combining expert knowledge and model optimization increased the excavation weights by up to 24.77% with low variance.