Stage M2/Ingénieur en robotique : Accelerating Task and Motion Planning through Geometric Reasoning

Context : Task and motion planning (TAMP) is a challenging robotics problem [1][2] which involves generating sequences of symbolic actions and their corresponding motions that enable a robot to achieve a desired goal. However, the primary challenge lies in addressing the high combinatorial complexity of combined task and motion planning. Traditional methods often require an extensive number of calls to motion planners to validate each planned action, making planning time-consuming and impractical for real-world applications. In our recent work [3][4][5], we leveraged deep learning methods to predict action and grasp feasibility. By doing so, we reduce the reliance on motion planners aiming to significantly accelerate the planning process.

Goal of the intern : The internship aims to extend this approach and improve its efficiency by providing the task and motion planner with additional geometric reasoning capabilities: ● Mapping the space of valid object placements to guide the currently random placement sampling. ● Leveraging the cause of action infeasibility to guide the search towards feasible solutions and improve the explainability of the planning process. ● Extending feasibility prediction to new robots (mobile manipulators) and problems with a heterogeneous set of robots. ● Considering optimality during the planning process.

The work will be structured in the following stages: ● Conduct a literature review on the existing TAMP methods and become familiar with the developed approach and the existing code. ● Investigate innovative methods to enhance the efficiency and augment the capabilities of the planning algorithms ● Evaluate the developed techniques in simulation and demonstrate experimentally the approach on real robots arms performing challenging TAMP problems.

Skills Required : ● Completing a Master’s program or an Engineering degree in Robotics, Computer Science, or related disciplines ● Good knowledge in robot modeling and planning ● Excellent programming skills, particularly in C++ and Python ● Good understanding of ROS ● Good knowledge of deep learning and some experience in task/motion planning are a plus

Duration and Location : The internship (4-6 months) will take place at LAAS-CNRS in Toulouse within the RIS team, providing the intern with a unique opportunity to work alongside experienced researchers and engineers in the field of robotics. The internship allowance is around 600 euros/month.

Application : Interested candidates are encouraged to submit their CV and a motivation letter to (saitbouhsa@laas.fr, simeon@laas.fr)

References

[1] Manipulation planning with probabilistic roadmaps. T. Siméon, J Cortés, JP. Laumond, A Sahbani. The International Journal of Robotics Research, (2004) (https://laas.hal.science/hal-01987879)

[2] A hybrid approach to intricate motion, manipulation and task planning. R. Alami, F. Gravot, S. Cambon. The International Journal of Robotics Research, (2009) (https://laas.hal.science/hal-01976081v1)

[3] Learning to predict action feasibility for task and motion planning in 3d environments. S. Ait Bouhsain, R. Alami, and T. Siméon. IEEE International Conference on Robotics and Automation (ICRA), 2023. (https://laas.hal.science/hal-03808885)

[4] Simultaneous Action and Grasp Feasibility Prediction for Task and Motion Planning through Multi-Task Learning. S. Ait Bouhsain, R. Alami, and T. Siméon, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2023. (https://laas.hal.science/hal-04016581)

[5] Extending Task and Motion Planning with Feasibility Prediction: Towards Multi-Robot Manipulation Planning of Realistic Objects. S. Ait Bouhsain, R. Alami, and T. Siméon. Technical Report, 2023. (https://hal.science/hal-04284213/)