Learning Feasible Transitions for Efficient Contact Planning
Akizhanov, Rikhat
Akizhanov, Rikhat
Author
Supervisor
Department
Computer Vision
Embargo End Date
2025-05-30
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
This project presents an efficient contact planning framework for quadrupedal robots navigating highly constrained environments, such as stepping stones. The challenge in such settings arises from the interplay between discrete decision making, selecting viable contact locations, and continuous trajectory optimization to ensure dynamic stability. The proposed approach in tegrates learning-based feasibility assessments with a structured search process, significantly improving computational efficiency and motion accuracy. Planning contact sequences for quadrupedal locomotion involves selecting foot placements while simultaneously optimizing centroidal dynamics and joint trajectories. Traditional methods often rely on exhaustive search techniques or manually defined heuristics, which are computationally expensive and lack scalability. To overcome these limitations, this project introduces a hybrid framework that combines Monte Carlo Tree Search (MCTS) with learned feasibility estimators and target adjustment mechanisms, facilitating rapid and accurate contact planning. A key contribution of this work is the development of a dynamic feasibility classifier, a neural network trained to predict whether a transition between two contact states is dynamically feasible. This classifier enables early pruning of infeasible transitions, reducing computational costs, and focusing search efforts on viable trajectories. Additionally, a target adjustment network is in troduced to compensate for low-level control inaccuracies by refining foot placements, ensuring precise execution despite controller imperfections. The proposed approach is implemented within an MCTS-based contact planner, where the learned networks are seamlessly integrated into the search process. Using offline training in di verse locomotion scenarios, the framework dynamically prunes infeasible transitions and refines contact placements, significantly accelerating planning while maintaining a high success rate. Extensive simulations validate the approach, demonstrating substantial reductions in search time and improvements in trajectory accuracy. Compared to baseline methods relying solely on kinematic feasibility checks or hand-crafted heuristics, the proposed method achieves superior performance in success rate, computational efficiency, and adaptability to varying terrain con ditions. Moreover, it generalizes well to unseen stepping stone configurations, reinforcing its robustness. The findings have significant implications for autonomous legged locomotion. By integrating data-driven heuristics within a structured search, the proposed approach bridges learningbased methods and classical planning techniques, offering a scalable solution for realworld deploy ment. Future work includes extending the framework to physical robots, optimizing gait se lection strategies, and exploring applications in loco-manipulation tasks. This research lays the foundation for more capable and autonomous quadrupedal robots in unstructured environments.
Citation
Rikhat Akizhanov, “Learning Feasible Transitions for Efficient Contact Planning,” Master of Science thesis, Computer Vision, MBZUAI, 2025.
Source
Conference
Keywords
Legged Robot, Contact Planning, Motion Planning, Deep Learning