RoboComp

Mentors: Student: Sushant Sreeram Swamy

Socially aware path planning enables a robot to navigate through a crowded environment causing the least amount of discomfort to the surrounding people. Building upon the work done in GSoC '21, the aim of the project is to account for the interactions between the different entities in the environment by adding interaction nodes between entities in the graph representation. A way in which the robot can also come up with interaction nodes would also be looked into.Finally, the project also aims to compare SNGNN-RL's performance with the current baseline.

Mentors: Student: Daniel Peix

Reinforcement learning is an established discipline for the automatic acquisition of control programs from a reward signal. Some early limitations arising from the necessary limited dimension of sensor spaces have been overcome with the integration of DNN as efficient reducers of large input data. In this project we want to apply RL as an online method to improve an existing grasping algorithm. Using a Kinova Gen3 arm simulated in CoppeliaSIm, that already performs a pick and place operation using a gripper with multiple force and distance sensors, the goal is to improve the current grasping performance by running RL algorithms that learn from the existing algorithm and, eventually, replace it to continue is adaptation to the specificities of the environment. If the experiments proceeds as expected, a second stage will try to transfer the learned controller to the grasping of a new target object in the scene, for example, from a block to a cup. All the experiments will be performed with RoboComp’s robotics cognitive architecture CORTEX, and the tools and existing agents already developed and tested.

Mentors: Student: Kaustab Pal

I am proposing the use of Convex Inner Approximations to plan the trajectories. The Convex Inner Approximation method finds kinodynamically feasible trajectories that guarantees collision avoidance. It also finds the trajectories in fewer iterations and as a result is much faster than traditional obstacle avoidance constraints.

Mentors: Student: Swati Dantu

Self-adaptation is a key ability for future autonomous robots. Thanks to this capability, a robot is able to automatically adapt to changes during its operation. In the case of autonomous navigation, if the robot is endowed with this skill, it can automatically adjust a set of free parameters to improve its functioning given a cost function or metric. The effect of this adaptation is safer, more efficient, and, possibly, better socially aware navigation. The aim of this project would be to design and implement a sliding variable based adaptive controller that can self-adapt under varying circumstances. Through this the robot would be able to successfully navigate and self adapt under social constraints as well if required.

Mentors: Student: Yassmin Hesham

<p>The field of Model Predictive Control (MPC) has seen tremendous progress. The algorithms and high-level software available to solve challenging nonlinear optimal control problems are significantly used in mobile robots to optimize in real-time their path following and navigation. In order to solve Non-Linear Programming Problems (NLP) – which is considered the general form –, we will use CasADi, an open-source tool to solve non-linear optimization problems. Also, the robot should be able to avoid obstacles in real-time by taking into consideration some constraints and penalizing the control values. In practice, previous maps and laser values are combined to create a list of convexified obstacle-free regions. To comply with the real-time requirements of a physical robot, the algorithm will be able to adapt – also in real-time –to the current situation by trading off between execution time and the number of constraints.</p>