8 December 2022 in IEEE Robotics and Automation Letters
In this letter, we present an approach for learning a neural implicit signed distance function expressed in joint space coordinates, that efficiently computes distance-to-collisions for arbitrary robotic manipulator configurations. Computing such distances is a long standing problem in robotics as approximate representations of the robot and environment geometry can lead to overly conservative constraints, numerical instabilities and expensive computations – limiting real-time reactive control and task success. Leveraging GPU parallelization and the differentiable nature of the proposed distance function allows for fast calculation of gradients with respect to the neural network inputs, providing a continuous repulsive vector field directly in joint space. We show that the learned high-resolution collision representation can be used to achieve real-time reactive control by i) formulating it as a collision-avoidance constraint for a quadratic programming (QP) inverse kinematics (IK), and ii) introducing it as a collision cost in a sampling-based joint space model predictive controller (MPC). For a reaching benchmark task with a 7DoF robot and dynamic obstacles intentionally obstructing the robot's path we achieve average 250 Hz control frequency with QP-IK and 92 Hz with MPC, showing an accelerated performance of 15% for QP-IK and 40% for MPC over baseline distance computation techniques.
3 February 2021 in IEEE Robotics and Automation Letters
In this letter, we propose a real-time self-collision avoidance approach for whole-body humanoid robot control. To achieve this, we learn the feasible regions of control in the humanoid's joint space as smooth self-collision boundary functions. Collision-free motions are generated online by treating the learned boundary functions as constraints in a Quadratic Program based Inverse Kinematic solver. As the geometrical complexity of a humanoid robot joint space grows with the number of degrees-of-freedom (DoF), learning computationally efficient and accurate boundary functions is challenging. We address this by partitioning the robot model into multiple lower-dimensional submodels. We compare performance of several state-of-the-art machine learning techniques to learn such boundary functions. Our approach is validated on the 29-DoF iCub humanoid robot, demonstrating highly accurate real-time self-collision avoidance.
12 April 2024 in The International Journal of Robotics Research
Dynamical system (DS) based motion planning offers collision-free motion, with closed-loop reactivity thanks to their analytical expression. It ensures that obstacles are not penetrated by reshaping a nominal DS through matrix modulation, which is constructed using continuously differentiable obstacle representations. However, state-of-the-art approaches may suffer from local minima induced by non-convex obstacles, thus failing to scale to complex, high-dimensional joint spaces. On the other hand, sampling-based Model Predictive Control (MPC) techniques provide feasible collision-free paths in joint-space, yet are limited to quasi-reactive scenarios due to computational complexity that grows cubically with space dimensionality and horizon length. To control the robot in the cluttered environment with moving obstacles, and to generate feasible and highly reactive collision-free motion in robots’ joint space, we present an approach for modulating joint-space DS using sampling-based MPC. Specifically, a nominal DS representing an unconstrained desired joint space motion to a target is locally deflected with obstacle-tangential velocity components navigating the robot around obstacles and avoiding local minima. Such tangential velocity components are constructed from receding horizon collision-free paths generated asynchronously by the sampling-based MPC. Notably, the MPC is not required to run constantly, but only activated when the local minima is detected. The approach is validated in simulation and real-world experiments on a 7-DoF robot demonstrating the capability of avoiding concave obstacles, while maintaining local attractor stability in both quasi-static and highly dynamic cluttered environments.
7 May 2018 in IEEE Robotics and Automation Letters
In this letter, we propose an asymptotically stable joint-space dynamical system (DS) that captures desired behaviors in joint-space while converging toward a task-space attractor in both position and orientation. To encode joint-space behaviors while meeting the stability criteria, we propose a DS constructed as a linear parameter varying system combining different behavior synergies and provide a method for learning these synergy matrices from demonstrations. Specifically, we use dimensionality reduction to find a low-dimensional embedding space for modulating joint synergies, and then estimate the parameters of the corresponding synergies by solving a convex semidefinite optimization problem that minimizes the joint velocity prediction error from the demonstrations. Our proposed approach is empirically validated on a variety of motions that reach a target in position and orientation, while following a desired joint-space behavior.
18 March 2024 in IEEE Transactions on Robotics
In this work, we present a novel, reactive, modulated control strategy based on dynamical systems (DS) for planning in the context of multiple non-convex obstacles. Our DS modulation strategy leverages an on-manifold planning methodology and provides several methods for real-time on-manifold navigation around non-convex obstacles. We introduce a sample-based obstacle representation for complex, non-convex obstacles, as well as a projection-based method for representing surfaces such as tables, cylinders, and ellipsoids. These representations can be combined to represent multiple obstacles and obstacle types (sample- or projection-based) with a single, continuously differentiable function. We validate our approach in several real-world scenarios, including navigation within (simulated) constrained environments, as well as reactive control of a real 7DoF manipulator with dynamic obstacles (including humans) while utilizing a 1 kHz control loop rate. Using our samplebased representation, we can calculate the obstacle representation function in less than 1 ms with up to 35k points using a CPU implementation, and up to 600k points with a GPU implementation.