Generative models for task and motion planning
| Název práce v češtině: | Generativní modely pro plánování úkolů a pohybů |
|---|---|
| Název v anglickém jazyce: | Generative models for task and motion planning |
| Klíčová slova: | plánování úkolů a pohybů|robotika|difuzní model |
| Klíčová slova anglicky: | task and motion planning|robotics|diffusion model |
| Akademický rok vypsání: | 2024/2025 |
| Typ práce: | diplomová práce |
| Jazyk práce: | angličtina |
| Ústav: | Katedra softwaru a výuky informatiky (32-KSVI) |
| Vedoucí / školitel: | Josef Šivic |
| Řešitel: | skrytý - zadáno a potvrzeno stud. odd. |
| Datum přihlášení: | 25.10.2024 |
| Datum zadání: | 30.10.2024 |
| Datum potvrzení stud. oddělením: | 31.10.2024 |
| Datum a čas obhajoby: | 04.02.2025 09:00 |
| Datum odevzdání elektronické podoby: | 09.01.2025 |
| Datum odevzdání tištěné podoby: | 09.01.2025 |
| Datum proběhlé obhajoby: | 04.02.2025 |
| Oponenti: | RNDr. David Obdržálek, Ph.D. |
| Konzultanti: | Ing. Vladimír Petrík, Ph.D. |
| Zásady pro vypracování |
| Task and motion planning is a crucial component of robotics that computes the motion of the robot and objects from the start configuration to the goal configuration while avoiding collisions and respecting the constraints of the robot and environment. However, the state-of-the-art task and motion planning algorithms are slow, taking minutes to plan a motion in a scene with two or more objects [1]. Recently, generative models were applied to solve robot motion planning [2, 3], i.e., for planning without objects. The goal of this thesis is to apply generative models for jointly planning robot motion with the motion of objects, i.e., task and motion planning. The training dataset will be created by solving task-and-motion planning problems using a classical RRT-based planning algorithm [4, 5, 6]. The goal is to study the generalization capability of the model to predict motions that were not seen during the training. More specifically, the objectives are:
1. Review state-of-the-art methods for using generative models for motion planning, such as [3]. Identify their limitations and benefits. 2. Generate training dataset for 2D task-and-motion planning by implementing RRT* Connect planner [6]. Ensure sufficient variability in the dataset by modifying obstacles, surfaces, objects, and start and goal configurations. 3. Design an architecture of a generative model and train it on the generated dataset. Analyze the generalization capabilities of the trained model by testing it in unseen scenarios. Quantitatively compare the generative planner with classical approaches (RRT Connect [5], RRT* Connect [6]). |
| Seznam odborné literatury |
| [1] K. Zorina et al., "Multi-Contact Task and Motion Planning Guided by Video Demonstration," 2023 IEEE International Conference on Robotics and Automation (ICRA), London, United Kingdom, 2023.
[2] Chaplot, D. S., Pathak, D., & Malik, J. (2021). Differentiable Spatial Planning using Transformers (Version 1). arXiv. [3] J. Carvalho et al., Motion Planning Diffusion: Learning and Planning of Robot Motions with Diffusion Models. IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2023. [4] LaValle, Steven M.. “Rapidly-exploring random trees: a new tool for path planning.” The annual research report, 1998. [5] J. J. Kuffner and S. M. LaValle, "RRT-connect: An efficient approach to single-query path planning," Proceedings 2000 ICRA. Millennium Conference. IEEE International Conference on Robotics and Automation. Symposia Proceedings (Cat. No.00CH37065), San Francisco, CA, USA, 2000. [6] S. Klemm et al., "RRT*-Connect: Faster, asymptotically optimal motion planning," 2015 IEEE International Conference on Robotics and Biomimetics (ROBIO), Zhuhai, China, 2015. [7] Yuan et al., PhysDiff: Physics-Guided Human Motion Diffusion Model. IEEE/CVF International Conference on Computer Vision (ICCV), 2023. |