We introduce GRiD: a GPU-accelerated library for computing rigid body dynamics with analytical gradients. GRiD was designed to accelerate the nonlinear trajectory opti-mization subproblem used in state-of-the-art robotic planning, control, and machine learning, which requires tens to hundreds of naturally parallel computations of rigid body dynamics and their gradients at each iteration. GRiD leverages URDF parsing and code generation to deliver optimized dynamics kernels that not only expose GPU-friendly computational patterns, but also take advantage of both fine-grained parallelism within each computation and coarse-grained parallelism between computations. Through this approach, when performing multiple computations of rigid body dynamics algorithms, GRiD provides as much as a 7.2x speedup over a state-of-the-art, multi-threaded CPU implementation, and maintains as much as a 2.5x speedup when accounting for I/O overhead. We release GRiD as an open-source library for use by the wider robotics community.

Recent human-in-the-loop (HIL) optimization studies using wearable devices have shown an improved average metabolic reduction by optimizing a small number of control parameters during short-duration walking experiments. However, the slow metabolic dynamics, high measurement noise, and experimental time constraints create challenges for increasing the number of control parameters to be optimized. Prior work applying gradient descent and Bayesian optimization to this problem have decoupled metabolic estimation and control parameter selection using fixed estimation intervals, which imposes a hard limit on the number of parameter evaluations possible in a given time budget. In this work, we take a different approach that couples estimation and parameter selection, allowing the algorithm to spend less time on refining the metabolic estimates for parameters that are unlikely to improve performance over the best observed values. Our approach uses a Kalman filter-based metabolic estimator to formulate an optimal stopping problem during the data acquisition step of standard Bayesian optimization. Performance was analyzed in numerical simulations and in pilot human subject testing with two subjects that involved optimizing six control parameters of a single-joint exosuit and four parameters of a multi-joint exosuit.

Designing reinforcement learning (RL) problems that can produce delicate and precise manipulation policies requires careful choice of the reward function, state, and action spaces. Much prior work on applying RL to manipulation tasks has defined the action space in terms of direct joint torques or reference positions for a joint-space proportional derivative (PD) controller. In practice, it is often possible to add additional structure by taking advantage of model-based controllers that support both accurate positioning and control of the dynamic response of the manipulator. In this paper, we evaluate how the choice of action space for dynamic manipulation tasks affects the sample complexity as well as the final quality of learned policies. We compare learning performance across three tasks (peg insertion, hammering, and pushing), four action spaces (torque, joint PD, inverse dynamics, and impedance control), and using two modern reinforcement learning algorithms (Proximal Policy optimization and Soft Actor-Critic). Our results lend support to the hypothesis that learning references for a task-space impedance controller significantly reduces the number of samples needed to achieve good performance across all tasks and algorithms.

Differential Dynamic Programming (DDP) has become a popular approach to performing trajectory optimization for complex, underactuated robots. However, DDP presents two practical challenges. First, the evaluation of dynamics derivatives during optimization creates a computational bottleneck, particularly in implementations that capture second-order dynamic effects. Second, constraints on the states (e.g., boundary conditions, collision constraints, etc.) require additional care since the state trajectory is implicitly defined from the inputs and dynamics. This paper addresses both of these problems by building on recent work on Unscented Dynamic Programming (UDP) - which eliminates dynamics derivative computations in DDP-to support general nonlinear state and input constraints using an augmented Lagrangian. The resulting algorithm has the same computational cost as first-order penalty-based DDP variants, but can achieve constraint satisfaction to high precision without the numerical ill-conditioning associated with penalty methods. We present results demonstrating its favorable performance on several simulated robot systems including a quadrotor and 7-DoF robot arm.

Contact constraints, such as those between a foot and the ground or a hand and an object, are inherent in many robotic tasks. These constraints define a manifold of feasible states; while well understood mathematically, they pose numerical challenges to many algorithms for planning and controlling whole-body dynamic motions. In this paper, we present an approach to the synthesis and stabilization of complex trajectories for both fully-actuated and underactuated robots subject to contact constraints. We introduce a trajectory optimization algorithm (DIRCON) that extends the direct collocation method, naturally incorporating manifold constraints to produce a nominal trajectory with third-order integration accuracy-a critical feature for achieving reliable tracking control. We adapt the classical time-varying linear quadratic regulator to produce a local cost-to-go in the manifold tangent plane. Finally, we descend the cost-to-go using a quadratic program that incorporates unilateral friction and torque constraints. This approach is demonstrated on three complex walking and climbing locomotion examples in simulation.

We describe a whole-body dynamic walking controller implemented as a convex quadratic program. The controller solves an optimal control problem using an approximate value function derived from a simple walking model while respecting the dynamic, input, and contact constraints of the full robot dynamics. By exploiting sparsity and temporal structure in the optimization with a custom active-set algorithm, we surpass the performance of the best available off-the-shelf solvers and achieve 1kHz control rates for a 34-DOF humanoid. We describe applications to balancing and walking tasks using the simulated Atlas robot in the DARPA Virtual Robotics Challenge.

The paper describes the system developed by researchers from MIT for the Defense Advanced Research Projects Agency's (DARPA) Virtual Robotics Challenge (VRC), held in June 2013. The VRC was the first competition in the DARPA Robotics Challenge (DRC), a program that aims to “develop ground robotic capabilities to execute complex tasks in dangerous, degraded, human-engineered environments”. The VRC required teams to guide a model of Boston Dynamics' humanoid robot, Atlas, through driving, walking, and manipulation tasks in simulation. Team MIT's user interface, the Viewer, provided the operator with a unified representation of all available information. A 3D rendering of the robot depicted its most recently estimated body state with respect to the surrounding environment, represented by point clouds and texture-mapped meshes as sensed by on-board LIDAR and fused over time.

We describe a robot system that autonomously acquires skills through interaction with its environment. The robot learns to sequence the execution of a set of innate controllers to solve a task, extracts and retains components of that solution as portable skills, and then transfers those skills to reduce the time required to learn to solve a second task.

The ability to discover the effects of actions and apply this knowledge during goal-oriented action selection is a fundamental requirement of embodied intelligent agents. In our ongoing work, we hope to demonstrate the utility of learned control models for whole-body mobile manipulation. In this short paper we discuss preliminary work on learning a forward model of the dynamics of a balancing robot exploring simple arm movements. This model is then used to construct whole-body control strategies for regulating state variables using arm motion.

We introduce CST, an algorithm for constructing skill trees from demonstration trajectories in continuous reinforcement learning domains. CST uses a changepoint detection method to segment each trajectory into a skill chain by detecting a change of appropriate abstraction, or that a segment is too complex to model as a single skill. The skill chains from each trajectory are then merged to form a skill tree. We demonstrate that CST constructs an appropriate skill tree that can be further refined through learning in a challenging continuous domain, and that it can be used to segment demonstration trajectories on a mobile manipulator into chains of skills where each skill is assigned an appropriate abstraction.