Achieving human-level performance on real world tasks is a north star for the robotics community. We present the first learned robot agent that reaches amateur humanlevel performance in competitive table tennis. Table tennis is a physically demanding sport that takes humans years to master. We contribute (1) a hierarchical and modular policy architecture consisting of (i) low level controllers with their skill descriptors that model their capabilities and (ii) a high level controller that chooses the low level skills, (2) techniques for enabling zero-shot sim-to-real and curriculum building, including an iterative approach (train in sim, deploy in real), and (3) real time adaptation to unseen opponents. Policy performance was assessed through 29 robot vs. human matches of which the robot won 45 % (13/29). All humans were unseen players and their skill level varied from beginner to tournament level. Whilst the robot lost all matches vs. the most advanced players it won 100 % matches vs. beginners and 55 % matches vs. intermediate players, demonstrating solidly amateur humanlevel performance. Videos of the matches can be viewed here1.See sites https://google.com/view/competitive-robot-table-tennis.

Contact-rich manipulation tasks with stiff frictional elements, like connector insertion, are difficult to model with rigid-body simulators. In this work, we propose a new approach for modeling these environments by learning a quasistatic contact force model instead of a full simulator. Using a feature vector that contains information about the configuration and control, we find a linear mapping adequately captures the relationship between this feature vector and the sensed contact forces. A novel Linear Model Learning (LML) algorithm is used to solve for the globally optimal mapping in real time without any matrix inversions, resulting in an algorithm that runs in nearly constant time on a GPU as the model size increases. We validate the proposed approach for connector insertion in both simulation and hardware experiments, where the learned model is combined with an optimizationbased impedance controller to achieve smooth insertions in the presence of misalignments and uncertainty. Our website featuring videos, code, and more materials is available at https://model-based-plugging.github.io/.

We address the longstanding challenge of producing flexible, realistic humanoid character controllers that can perform diverse whole-body tasks involving object interactions. This challenge is central to a variety of fields, from graphics and animation to robotics and motor neuroscience. Our physics-based environment uses realistic actuation and first-person perception - including touch sensors and egocentric vision - with a view to producing active-sensing behaviors (e.g. gaze direction), transferability to real robots, and comparisons to the biology. We develop an integrated neural-network based approach consisting of a motor primitive module, human demonstrations, and an instructed reinforcement learning regime with curricula and task variations. We demonstrate the utility of our approach for several tasks, including goal-conditioned box carrying and ball catching, and we characterize its behavioral robustness. The resulting controllers can be deployed in real-time on a standard PC. 1

We present a framework for efficient inference in structured image models that explicitly reason about objects. We achieve this by performing probabilistic inference using a recurrent neural network that attends to scene elements and processes them one at a time. Crucially, the model itself learns to choose the appropriate number of inference steps. We use this scheme to learn to perform inference in partially specified 2D models (variable-sized variational auto-encoders) and fully specified 3D models (probabilistic renderers). We show that such models learn to identify multiple objects - counting, locating and classifying the elements of a scene - without any supervision, e.g., decomposing 3D images with various numbers of objects in a single forward pass of a neural network at unprecedented speed. We further show that the networks produce accurate inferences when compared to supervised counterparts, and that their structure leads to improved generalization.

There is growing need for software tools that can accurately simulate the complex dynamics of modern robots. While a number of candidates exist, the field is fragmented. It is difficult to select the best tool for a given project, or to predict how much effort will be needed and what the ultimate simulation performance will be. Here we introduce new quantitative measures of simulation performance, focusing on the numerical challenges that are typical for robotics as opposed to multi-body dynamics and gaming. We then present extensive simulation results, obtained within a new software framework for instantiating the same model in multiple engines and running side-by-side comparisons. Overall we find that each engine performs best on the type of system it was designed and optimized for: MuJoCo wins the robotics-related tests, while the gaming engines win the gaming-related tests without a clear leader among them. The simulations are illustrated in the accompanying movie.

Controlling the robot with a permanently-updated optimal trajectory, also known as model predictive control, is the Holy Grail of whole-body motion generation. Before obtaining it, several challenges should be faced: computation cost, non-linear local minima, algorithm stability, etc. In this paper, we address the problem of applying the updated optimal control in real-time on the physical robot. In particular, we focus on the problems raised by the delays due to computation and by the differences between the real robot and the simulated model. Based on the optimal-control solver MuJoCo, we implemented a complete model-predictive controller and we applied it in real-time on the physical HRP-2 robot. It is the first time that such a whole-body model predictive controller is applied in real-time on a complex dynamic robot. Aside from the technical contributions cited above, the main contribution of this paper is to report the experimental results of this première implementation.

We present a unified framework for learning continuous control policies usingbackpropagation. It supports stochastic control by treating stochasticity in theBellman equation as a deterministic function of exogenous noise. The productis a spectrum of general policy gradient algorithms that range from model-freemethods with value functions to model-based methods without value functions.We use learned models but only require observations from the environment insteadof observations from model-predicted trajectories, minimizing the impactof compounded model errors. We apply these algorithms first to a toy stochasticcontrol problem and then to several physics-based control problems in simulation.One of these variants, SVG(1), shows the effectiveness of learning models, valuefunctions, and policies simultaneously in continuous domains.

Dexterous hand manipulation is one of the most complex types of biological movement, and has proven very difficult to replicate in robots. The usual approaches to robotic control - following pre-defined trajectories or planning online with reduced models - are both inapplicable. Dexterous manipulation is so sensitive to small variations in contact force and object location that it seems to require online planning without any simplifications. Here we demonstrate for the first time online planning (or model-predictive control) with a full physics model of a humanoid hand, with 28 degrees of freedom and 48 pneumatic actuators. We augment the actuation space with motor synergies which speed up optimization without removing dexterity. Most of our results are in simulation, showing non-prehensile object manipulation as well as typing. In both cases the input to the system is a high level task description, while all details of the hand movement emerge online from fully automated numerical optimization.

Trajectory optimizers are a powerful class of methods for generating goal-directed robot motion. Differential Dynamic Programming (DDP) is an indirect method which optimizes only over the unconstrained control-space and is therefore fast enough to allow real-time control of a full humanoid robot on modern computers. Although indirect methods automatically take into account state constraints, control limits pose a difficulty. This is particularly problematic when an expensive robot is strong enough to break itself. In this paper, we demonstrate that simple heuristics used to enforce limits (clamping and penalizing) are not efficient in general. We then propose a generalization of DDP which accommodates box inequality constraints on the controls, without significantly sacrificing convergence quality or computational effort. We apply our algorithm to three simulated problems, including the 36-DoF HRP-2 robot. A movie of our results can be found here goo.gl/eeiMnn.

We describe an accurate approach to state estimation which fuses any available sensor data with physical consistency priors. This is done by combining the advantages of recursive estimation and fixed-lag smoothing: at each step we re-estimate the trajectory over a time window into the past, but also use a recursive prior obtained from the previous time step via internal simulation. We also incorporate a physics engine into the estimator, which makes it possible to adjust the state estimates so that the inferred contact interactions are consistent with the observed accelerations. The estimator can utilize contact sensors to improve accuracy, but even in the absence of such sensors it reasons correctly about contact forces. Estimation speed and accuracy are demonstrated on a 28-DOF humanoid robot (Darwin) in a walking task. Timing tests and leave-one-out cross-validation show that the proposed approach can be used in real-time and is substantially more accurate than the EKF, without any over-fitting. A video of our results is attached.

Both global methods and on-line trajectory optimization methods are powerful techniques for solving optimal control problems; however, each has limitations. In order to mitigate the undesirable properties of each, we explore the possibility of combining the two. We explore two methods of deriving a descriptive final cost function to assist model predictive control (MPC) in selecting a good policy without having to plan as far into the future or having to fine-tune delicate cost functions. First, we exploit the large amount of data which is generated in MPC simulations (based on the receding horizon iterative LQG method) to learn, off-line, the global optimal value function for use as a final cost. We demonstrate that, while the global function approximation matches the value function well on some problems, there is relatively little improvement to the original MPC. Alternatively, we solve the Bellman equation directly using aggregation methods for linearly-solvable Markov Decision Processes to obtain an approximation to the value function and the optimal policy. Using both pieces of information in the MPC framework, we find controller performance of similar quality to MPC alone with long horizon, but now we may drastically shorten the horizon. Implementation of these methods shows that Bellman equation-based methods and on-line trajectory methods can be combined in real applications to the benefit of both.

We present an online trajectory optimization method and software platform applicable to complex humanoid robots performing challenging tasks such as getting up from an arbitrary pose on the ground and recovering from large disturbances using dexterous acrobatic maneuvers. The resulting behaviors, illustrated in the attached video, are computed only 7 × slower than real time, on a standard PC. The video also shows results on the acrobot problem, planar swimming and one-legged hopping. These simpler problems can already be solved in real time, without pre-computing anything.

We describe a new physics engine tailored to model-based control. Multi-joint dynamics are represented in generalized coordinates and computed via recursive algorithms. Contact responses are computed via efficient new algorithms we have developed, based on the modern velocity-stepping approach which avoids the difficulties with spring-dampers. Models are specified using either a high-level C++ API or an intuitive XML file format. A built-in compiler transforms the user model into an optimized data structure used for runtime computation. The engine can compute both forward and inverse dynamics. The latter are well-defined even in the presence of contacts and equality constraints. The model can include tendon wrapping as well as actuator activation states (e.g. pneumatic cylinders or muscles). To facilitate optimal control applications and in particular sampling and finite differencing, the dynamics can be evaluated for different states and controls in parallel. Around 400,000 dynamics evaluations per second are possible on a 12-core machine, for a 3D homanoid with 18 dofs and 6 active contacts. We have already used the engine in a number of control applications. It will soon be made publicly available.

We describe a new local dynamic programming algorithm for solving stochastic continuous Optimal Control problems. We use cubature integration to both propagate the state distribution and perform the Bellman backup. The algorithm can approximate the local policy and cost-to-go with arbitrary function bases. We compare the classic quadratic cost-to-go/linear-feedback controller to a cubic cost-to-go/quadratic policy controller on a 10-dimensional simulated swimming robot, and find that the higher order approximation yields a more general policy with a larger basin of attraction.

We develop an iterative local dynamic programming method (iLDP) applicable to stochastic optimal control problems in continuous high-dimensional state and action spaces. Such problems are common in the control of biological movement, but cannot be handled by existing methods. iLDP can be considered a generalization of differential dynamic programming, in as much as: (a) we use general basis functions rather than quadratics to approximate the optimal value function; (b) we introduce a collocation method that dispenses with explicit differentiation of the cost and dynamics and ties iLDP to the unscented Kalman filter; (c) we adapt the local function approximator to the propagated state covariance, thus increasing accuracy at more likely states. Convergence is similar to quasi-Newton methods. We illustrate iLDP on several problems including the ldquoswimmerrdquo dynamical system which has 14 state and 4 control variables.

The control of high-dimensional, continuous, non-linear systems is a key problem in reinforcement learning and control. Local, trajectory-based methods, using techniques such as Differential Dynamic Programming (DDP) are not directly subject to the curse of dimensionality, but generate only local controllers. In this paper, we introduce Receding Horizon DDP (RH-DDP), an extension to the classic DDP algorithm, which allows us to construct stable and robust controllers based on a library of local-control trajectories. We demonstrate the effectiveness of our approach on a series of high-dimensional control problems using a simulated multi-link swimming robot. These experiments show that our approach effectively circumvents dimensionality issues, and is capable of dealing effectively with problems with (at least) 34 state and 14 action dimensions.

In this paper, we present an empirical study of iterative least squares minimization of the Hamilton-Jacobi-Bellman (HJB) residual with a neural network (NN) approximation of the value function. Although the nonlinearities in the optimal control problem and NN approximator preclude theoretical guarantees and raise concerns of numerical instabilities, we present two simple methods for promoting convergence, the effectiveness of which is presented in a series of experiments. The first method involves the gradual increase of the horizon time scale, with a corresponding gradual increase in value function complexity. The second method involves the assumption of stochastic dynamics which introduces a regularizing second derivative term to the HJB equation. A gradual reduction of this term provides further stabilization of the convergence. We demonstrate the solution of several problems, including the 4-D inverted-pendulum system with bounded control. Our approach requires no initial stabilizing policy or any restrictive assumptions on the plant or cost function, only knowledge of the plant dynamics. In the Appendix, we provide the equations for first- and second-order differential backpropagation.