We present DemoStart, a novel auto-curriculum reinforcement learning method capable of learning complex manipulation behaviors on an arm equipped with a three- fingered robotic hand, from only a sparse reward and a handful of demonstrations in simulation. Learning from simulation drastically reduces the development cycle of behavior generation, and domain randomization techniques are leveraged to achieve successful zero-shot sim-to- real transfer. Transferred policies are learned directly from raw pixels from multiple cameras and robot proprioception. Our approach outperforms policies learned from demonstrations on the real robot and requires 100 times fewer demonstrations, collected in simulation. More details and videos in sites.google.com/view/demostart.

Modern robotic manufacturing requires collision-free coordination of multiple robots to complete numerous tasks in shared, obstacle-rich workspaces. Although individual tasks may be simple in isolation, automated joint task allocation, scheduling, and motion planning under spatiotemporal constraints remain computationally intractable for classical methods at real-world scales. Existing multiarm systems deployed in industry rely on human intuition and experience to design feasible trajectories manually in a labor-intensive process. To address this challenge, we propose a reinforcement learning (RL) framework to achieve automated task and motion planning, tested in an obstacle-rich environment with eight robots performing 40 reaching tasks in a shared workspace, where any robot can perform any task in any order. Our approach builds on a graph neural network (GNN) policy trained via RL on procedurally generated environments with diverse obstacle layouts, robot configurations, and task distributions. It uses a graph representation of scenes and a graph policy neural network trained through RL to generate trajectories of multiple robots, jointly solving the subproblems of task allocation, scheduling, and motion planning. Trained on large randomly generated task sets in simulation, our policy generalizes zero-shot to unseen settings with varying robot placements, obstacle geometries, and task poses. We further demonstrate that the high-speed capability of our solution enables its use in workcell layout optimization, improving solution times. The speed and scalability of our planner also open the door to capabilities such as fault-tolerant planning and online perception-based replanning, where rapid adaptation to dynamic task sets is required.

The ability to leverage heterogeneous robotic experience from different robots and tasks to quickly master novel skills and embodiments has the potential to transform robot learning. Inspired by recent advances in foundation models for vision and language, we propose a multi-embodiment, multi-task generalist agent for robotic manipulation. This agent, named RoboCat, is a visual goal-conditioned decision transformer capable of consuming action-labelled visual experience. This data spans a large repertoire of motor control skills from simulated and real robotic arms with varying sets of observations and actions. With RoboCat, we demonstrate the ability to generalise to new tasks and robots, both zero-shot as well as through adaptation using only 100-1000 examples for the target task. We also show how a trained model itself can be used to generate data for subsequent training iterations, thus providing a basic building block for an autonomous improvement loop. We investigate the agent's capabilities, with large-scale evaluations both in simulation and on three different real robot embodiments. We find that as we grow and diversify its training data, RoboCat not only shows signs of cross-task transfer, but also becomes more efficient at adapting to new tasks.

Many robot manipulation tasks can be framed as geometric reasoning tasks, where an agent must be able to precisely manipulate an object into a position that satisfies the task from a set of initial conditions. Often, task success is defined based on the relationship between two objects - for instance, hanging a mug on a rack. In such cases, the solution should be equivariant to the initial position of the objects as well as the agent, and invariant to the pose of the camera. This poses a challenge for learning systems which attempt to solve this task by learning directly from high-dimensional demonstrations: the agent must learn to be both equivariant as well as precise, which can be challenging without any inductive biases about the problem. In this work, we propose a method for precise relative pose prediction which is provably SE(3)-equivariant, can be learned from only a few demonstrations, and can generalize across variations in a class of objects. We accomplish this by factoring the problem into learning an SE(3) invariant task-specific representation of the scene and then interpreting this representation with novel geometric reasoning layers which are provably SE(3) equivariant. We demonstrate that our method can yield substantially more precise placement predictions in simulated placement tasks than previous methods trained with the same amount of data, and can accurately represent relative placement relationships data collected from real-world demonstrations. Supplementary information and videos can be found at https://sites.google.com/view/reldist-iclr-2023.

For robots to be useful outside labs and specialized factories we need a way to teach them new useful behaviors quickly. Current approaches lack either the generality to onboard new tasks without task-specific engineering, or else lack the data-efficiency to do so in an amount of time that enables practical use. In this work we explore dense tracking as a representational vehicle to allow faster and more general learning from demonstration. Our approach utilizes Track-Any-Point (TAP) models to isolate the relevant motion in a demonstration, and parameterize a low-level controller to reproduce this motion across changes in the scene configuration. We show this results in robust robot policies that can solve complex object-arrangement tasks such as shape-matching, stacking, and even full path-following tasks such as applying glue and sticking objects together, all from demonstrations that can be collected in minutes.

A robot's ability to act is fundamentally constrained by what it can perceive. Many existing approaches to visual representation learning utilize general-purpose training criteria, e.g. image reconstruction, smoothness in latent space, or usefulness for control, or else make use of large datasets annotated with specific features (bounding boxes, segmentations, etc.). However, both approaches often struggle to capture the fine-detail required for precision tasks on specific objects, e.g. grasping and mating a plug and socket. We argue that these difficulties arise from a lack of geometric structure in these models. In this work we advocate semantic 3D keypoints as a visual representation, and present a semi-supervised training objective that can allow instance or category-level keypoints to be trained to 1-5 millimeter-accuracy with minimal supervision. Furthermore, unlike local texture-based approaches, our model integrates contextual information from a large area and is therefore robust to occlusion, noise, and lack of discernible texture. We demonstrate that this ability to locate semantic keypoints enables high level scripting of human understandable behaviours. Finally we show that these keypoints provide a good way to define reward functions for reinforcement learning and are a good representation for training agents.

We present a framework for data-driven robotics that makes use of a large dataset of recorded robot experience and scales to several tasks using learned reward functions. We show how to apply this framework to accomplish three different object manipulation tasks on a real robot platform. Given demonstrations of a task together with task-agnostic recorded experience, we use a special form of human annotation as supervision to learn a reward function, which enables us to deal with real-world tasks where the reward signal cannot be acquired directly. Learned rewards are used in combination with a large dataset of experience from different tasks to learn a robot policy offline using batch RL. We show that using our approach it is possible to train agents to perform a variety of challenging manipulation tasks including stacking rigid objects and handling cloth.

We propose Generative Predecessor Models for Imitation Learning (GPRIL), a novel imitation learning algorithm that matches the state-action distribution to the distribution observed in expert demonstrations, using generative models to reason probabilistically about alternative histories of demonstrated states. We show that this approach allows an agent to learn robust policies using only a small number of expert demonstrations and self-supervised interactions with the environment. We derive this approach from first principles and compare it empirically to a state-of-the-art imitation learning method, showing that it outperforms or matches its performance on two simulated robot manipulation tasks and demonstrate significantly higher sample efficiency by applying the algorithm on a real robot.

We present a framework for data-driven robotics that makes use of a large dataset of recorded robot experience and scales to several tasks using learned reward functions. We show how to apply this framework to accomplish three different object manipulation tasks on a real robot platform. Given demonstrations of a task together with task-agnostic recorded experience, we use a special form of human annotation as supervision to learn a reward function, which enables us to deal with real-world tasks where the reward signal cannot be acquired directly. Learned rewards are used in combination with a large dataset of experience from different tasks to learn a robot policy offline using batch RL. We show that using our approach it is possible to train agents to perform a variety of challenging manipulation tasks including stacking rigid objects and handling cloth.

Model-free reinforcement learning algorithms such as Deep Deterministic Policy Gradient (DDPG) often require additional exploration strategies, especially if the actor is of deterministic nature. This work evaluates the use of model-based trajectory optimization methods used for exploration in Deep Deterministic Policy Gradient when trained on a latent image embedding. In addition, an extension of DDPG is derived using a value function as critic, making use of a learned deep dynamics model to compute the policy gradient. This approach leads to a symbiotic relationship between the deep reinforcement learning algorithm and the latent trajectory optimizer. The trajectory optimizer benefits from the critic learned by the RL algorithm and the latter from the enhanced exploration generated by the planner. The developed methods are evaluated on two continuous control tasks, one in simulation and one in the real world. In particular, a Baxter robot is trained to perform an insertion task, while only receiving sparse rewards and images as observations from the environment.

We propose position-velocity encoders (PVEs) which learn---without supervision---to encode images to positions and velocities of task-relevant objects. PVEs encode a single image into a low-dimensional position state and compute the velocity state from finite differences in position. In contrast to autoencoders, position-velocity encoders are not trained by image reconstruction, but by making the position-velocity representation consistent with priors about interacting with the physical world. We applied PVEs to several simulated control tasks from pixels and achieved promising preliminary results.

We propose position-velocity encoders (PVEs) which learn—without supervision—to encode images to positions and velocities of task-relevant objects. PVEs encode a single image into a low-dimensional position state and compute the velocity state from finite differences in position. In contrast to autoencoders, position-velocity encoders are not trained by image reconstruction, but by making the position-velocity representation consistent with priors about interacting with the physical world. We applied PVEs to several simulated control tasks from pixels and achieved promising preliminary results.

We propose a general and model-free approach for Reinforcement Learning (RL) on real robotics with sparse rewards. We build upon the Deep Deterministic Policy Gradient (DDPG) algorithm to use demonstrations. Both demonstrations and actual interactions are used to fill a replay buffer and the sampling ratio between demonstrations and transitions is automatically tuned via a prioritized replay mechanism. Typically, carefully engineered shaping rewards are required to enable the agents to efficiently explore on high dimensional control problems such as robotics. They are also required for model-based acceleration methods relying on local solvers such as iLQG (e.g. Guided Policy Search and Normalized Advantage Function). The demonstrations replace the need for carefully engineered rewards, and reduce the exploration problem encountered by classical RL approaches in these domains. Demonstrations are collected by a robot kinesthetically force-controlled by a human demonstrator. Results on four simulated insertion tasks show that DDPG from demonstrations out-performs DDPG, and does not require engineered rewards. Finally, we demonstrate the method on a real robotics task consisting of inserting a clip (flexible object) into a rigid object.

In this paper we present the first planner for the problem of Navigation Among Movable Obstacles (NAMO) on a real robot that can handle environments with under-specified object dynamics. This result makes use of recent progress from two threads of the Reinforcement Learning literature. The first is a hierarchical Markov-Decision Process formulation of the NAMO problem designed to handle dynamics uncertainty. The second is a physics-based Reinforcement Learning framework which offers a way to ground this uncertainty in a compact model space that can be efficiently updated from data received by the robot online. Our results demonstrate the ability of a robot to adapt to unexpected object behavior in a real office scenario.

For a mobile manipulator to interact with large everyday objects, such as office tables, it is often important to have dynamic models of these objects. However, as it is infeasible to provide the robot with models for every possible object it may encounter, it is desirable that the robot can identify common object models autonomously. Existing methods for addressing this challenge are limited by being either purely kinematic, or inefficient due to a lack of physical structure. In this paper, we present a physics-based method for estimating the dynamics of common non-holonomic objects using a mobile manipulator, and demonstrate its efficiency compared to existing approaches.

One of the key challenges in using reinforcement learning in robotics is the need for models that capture natural world structure. There are methods that formalize multi-object dynamics using relational representations, but these methods are not sufficiently compact for real-world robotics. We present a physics-based approach that exploits modern simulation tools to efficiently parameterize physical dynamics. Our results show that this representation can result in much faster learning, by virtue of its strong but appropriate inductive bias in physical environments.

In this paper we present a decision theoretic planner for the problem of Navigation Among Movable Obstacles (NAMO) operating under conditions faced by real robotic systems. While planners for the NAMO domain exist, they typically assume a deterministic environment or rely on discretization of the configuration and action spaces, preventing their use in practice. In contrast, we propose a planner that operates in real-world conditions such as uncertainty about the parameters of workspace objects and continuous configuration and action (control) spaces.To achieve robust NAMO planning despite these conditions, we introduce a novel integration of Monte Carlo simulation with an abstract MDP construction. We present theoretical and empirical arguments for time complexity linear in the number of obstacles as well as a detailed implementation and examples from a dynamic simulation environment.

In this paper we present the first decision theoretic planner for the problem of Navigation Among Movable Obstacles (NAMO). While efficient planners for NAMO exist, they are challenging to implement in practice due to the inherent uncertainty in both perception and control of real robots. Generalizing existing NAMO planners to nondeterministic domains is particularly difficult due to the sensitivity of MDP methods to task dimensionality. Our work addresses this challenge by combining ideas from Hierarchical Reinforcement Learning with Monte Carlo Tree Search, and results in an algorithm that can be used for fast online planning in uncertain environments. We evaluate our algorithm in simulation, and provide a theoretical argument for our results which suggest linear time complexity in the number of obstacles for typical environments.

We explore robot object manipulation as a Bayesian model-based reinforcement learning problem under a collection of different model priors. Our main contribution is to highlight the limitations of classical non-parametric regression approaches in the context of online learning, and to introduce an alternative approach based on monolithic physical inference. The primary motivation for this line of research is to incorporate physical system identification into the RL model, where it can be integrated with modern approaches to Bayesian structure learning. Overall, our results support the idea that modern physical simulation tools provide a model space with an appropriate inductive bias for manipulation problems in natural environments.