CoRL 2022 Papers

Learned locomotion policies can rapidly adapt to diverse environments similar to those experienced during training but lack a mechanism for fast tuning when they fail in an out-of-distribution test environment. This necessitates a slow and iterative cycle of reward and environment redesign to achieve good performance on a new task. As an alternative, we propose learning a single policy that encodes a structured family of locomotion strategies that solve training tasks in different ways, resulting in Multiplicity of Behavior (MoB). Different strategies generalize differently and can be chosen in real-time for new tasks or environments, bypassing the need for time-consuming retraining. We release a fast, robust open-source MoB locomotion controller, Walk These Ways, that can execute diverse gaits with variable footswing, posture, and speed, unlocking diverse downstream tasks: crouching, hopping, high-speed running, stair traversal, bracing against shoves, rhythmic dance, and more. Video and code release: https://gmargo11.github.io/walk-these-ways

Sim-to-real transfer is a powerful paradigm for robotic reinforcement learning. The ability to train policies in simulation enables safe exploration and large-scale data collection quickly at low cost. However, prior works in sim-to-real transfer of robotic policies typically do not involve any human-robot interaction because accurately simulating human behavior is an open problem. In this work, our goal is to leverage the power of simulation to train robotic policies that are proficient at interacting with humans upon deployment. But there is a chicken and egg problem --- how to gather examples of a human interacting with a physical robot so as to model human behavior in simulation without already having a robot that is able to interact with a human? Our proposed method, Iterative-Sim-to-Real (i-S2R), attempts to address this. i-S2R bootstraps from a simple model of human behavior and alternates between training in simulation and deploying in the real world. In each iteration, both the human behavior model and the policy are refined. For all training we apply a new evolutionary search algorithm called Blackbox Gradient Sensing (BGS). We evaluate our method on a real world robotic table tennis setting, where the objective for the robot is to play cooperatively with a human player for as long as possible. Table tennis is a high-speed, dynamic task that requires the two players to react quickly to each other’s moves, making for a challenging test bed for research on human-robot interaction. We present results on an industrial robotic arm that is able to cooperatively play table tennis with human players, achieving rallies of 22 successive hits on average and 150 at best. Further, for 80% of players, rally lengths are 70% to 175% longer compared to the sim-to-real plus fine-tuning (S2R+FT) baseline. For videos of our system in action please see https://sites.google.com/view/is2r.

For tabletop object manipulation tasks, learning an accurate pushing dynamics model, which predicts the objects' motions when a robot pushes an object, is very important. In this work, we claim that an ideal pushing dynamics model should have the SE(2)-equivariance property, i.e., if tabletop objects' poses and pushing action are transformed by some same planar rigid-body transformation, then the resulting motion should also be the result of the same transformation. Existing state-of-the-art data-driven approaches do not have this equivariance property, resulting in less-than-desirable learning performances. In this paper, we propose a new neural network architecture that by construction has the above equivariance property. Through extensive empirical validations, we show that the proposed model shows significantly improved learning performances over the existing methods. Also, we verify that our pushing dynamics model can be used for various downstream pushing manipulation tasks such as the object moving, singulation, and grasping in both simulation and real robot experiments. Code is available at https://github.com/seungyeon-k/SQPDNet-public.

We propose new methods to synthesize control barrier function (CBF) based safe controllers that avoid input saturation, which can cause safety violations. In particular, our method is created for high-dimensional, general nonlinear systems, for which such tools are scarce. We leverage techniques from machine learning, like neural networks and deep learning, to simplify this challenging problem in nonlinear control design. The method consists of a learner-critic architecture, in which the critic gives counterexamples of input saturation and the learner optimizes a neural CBF to eliminate those counterexamples. We provide empirical results on a 10D state, 4D input quadcopter-pendulum system. Our learned CBF avoids input saturation and maintains safety over nearly 100% of trials.

To solve tasks in complex environments, robots need to learn from experience. Deep reinforcement learning is a common approach to robot learning but requires a large amount of trial and error to learn, limiting its deployment in the physical world. As a consequence, many advances in robot learning rely on simulators. On the other hand, learning inside of simulators fails to capture the complexity of the real world, is prone to simulator inaccuracies, and the resulting behaviors do not adapt to changes in the world. The Dreamer algorithm has recently shown great promise for learning from small amounts of interaction by planning within a learned world model, outperforming pure reinforcement learning in video games. Learning a world model to predict the outcomes of potential actions enables planning in imagination, reducing the amount of trial and error needed in the real environment. However, it is unknown whether Dreamer can facilitate faster learning on physical robots. In this paper, we apply Dreamer to 4 robots to learn online and directly in the real world, without any simulators. Dreamer trains a quadruped robot to roll off its back, stand up, and walk from scratch and without resets in only 1 hour. We then push the robot and find that Dreamer adapts within 10 minutes to withstand perturbations or quickly roll over and stand back up. On two different robotic arms, Dreamer learns to pick and place objects from camera images and sparse rewards, approaching human-level teleoperation performance. On a wheeled robot, Dreamer learns to navigate to a goal position purely from camera images, automatically resolving ambiguity about the robot orientation. Using the same hyperparameters across all experiments, we find that Dreamer is capable of online learning in the real world, which establishes a strong baseline. We release our infrastructure for future applications of world models to robot learning.

Recent works have shown how the reasoning capabilities of Large Language Models (LLMs) can be applied to domains beyond natural language processing, such as planning and interaction for robots. These embodied problems require an agent to understand many semantic aspects of the world: the repertoire of skills available, how these skills influence the world, and how changes to the world map back to the language. LLMs planning in embodied environments need to consider not just what skills to do, but also how and when to do them - answers that change over time in response to the agent's own choices. In this work, we investigate to what extent LLMs used in such embodied contexts can reason over sources of feedback provided through natural language, without any additional training. We propose that by leveraging environment feedback, LLMs are able to form an inner monologue that allows them to more richly process and plan in robotic control scenarios. We investigate a variety of sources of feedback, such as success detection, scene description, and human interaction. We find that closed-loop language feedback significantly improves high level instruction completion on three domains, including simulated and real table top rearrangement tasks and long-horizon mobile manipulation tasks in a kitchen environment in the real world.

Research on Interactive Robot Learning has yielded several modalities for querying a human for training data, including demonstrations, preferences, and corrections. While prior work in this space has focused on optimizing the robot's queries within each interaction type, there has been little work on optimizing over the selection of the interaction type itself. We present INQUIRE, the first algorithm to implement and optimize over a generalized representation of information gain across multiple interaction types. Our evaluations show that INQUIRE can dynamically optimize its interaction type (and respective optimal query) based on its current learning status and the robot's state in the world, resulting in more robust performance across tasks in comparison to state-of-the art baseline methods. Additionally, INQUIRE allows for customizable cost metrics to bias its selection of interaction types, enabling this algorithm to be tailored to a robot's particular deployment domain and formulate cost-aware, informative queries.

Point clouds are a widely available and canonical data modality which convey the 3D geometry of a scene. Despite significant progress in classification and segmentation from point clouds, policy learning from such a modality remains challenging, and most prior works in imitation learning focus on learning policies from images or state information. In this paper, we propose a novel framework for learning policies from point clouds for robotic manipulation with tools. We use a novel neural network, ToolFlowNet, which predicts dense per-point flow on the tool that the robot controls, and then uses the flow to derive the transformation that the robot should execute. We apply this framework to imitation learning of challenging deformable object manipulation tasks with continuous movement of tools, including scooping and pouring, and demonstrate significantly improved performance over baselines which do not use flow. We perform physical scooping experiments with ToolFlowNet and find that we can attain 82% scooping success. See https://sites.google.com/view/point-cloud-policy/home for supplementary material.

We propose an image-based, learned method for selective tabletop object rearrangement in clutter using a parallel jaw gripper. Our method consists of three stages: graph-based object sequencing (which object to move), feature-based action selection (whether to push or grasp, and at what position and orientation) and a visual correspondence-based placement policy (where to place a grasped object). Experiments show that this decomposition works well in challenging settings requiring the robot to begin with an initially cluttered scene, selecting only the objects that need to be rearranged while discarding others, and dealing with cases where the goal location for an object is already occupied – making it the first system to address all these concurrently in a purely image-based setting. We also achieve an $\sim$ 8% improvement in task success rate over the previously best reported result that handles both translation and orientation in less restrictive (un-cluttered, non-selective) settings. We demonstrate zero-shot transfer of our system solely trained in simulation to a real robot selectively rearranging up to everyday objects, many unseen during learning, on a crowded tabletop. Videos:https://sites.google.com/view/selective-rearrangement

Robots in complex multi-agent environments should reason about the intentions of observed and currently unobserved agents. In this paper, we present a new learning-based method for prediction and planning in complex multi-agent environments where the states of the other agents are partially-observed. Our approach, Active Visual Planning (AVP), uses high-dimensional observations to learn a flow-based generative model of multi-agent joint trajectories, including unobserved agents that may be revealed in the near future, depending on the robot's actions. Our predictive model is implemented using deep neural networks that map raw observations to future detection and pose trajectories and is learned entirely offline using a dataset of recorded observations (not ground-truth states). Once learned, our predictive model can be used for contingency planning over the potential existence, intentions, and positions of unobserved agents. We demonstrate the effectiveness of AVP on a set of autonomous driving environments inspired by real-world scenarios that require reasoning about the existence of other unobserved agents for safe and efficient driving. In these environments, AVP achieves optimal closed-loop performance, while methods that do not reason about potential unobserved agents exhibit either overconfident or underconfident behavior.

We introduce MegaPose, a method to estimate the 6D pose of novel objects, that is, objects unseen during training. At inference time, the method only assumes knowledge of (i) a region of interest displaying the object in the image and (ii) a CAD model of the observed object. The contributions of this work are threefold. First, we present a 6D pose refiner based on a render&compare strategy which can be applied to novel objects. The shape and coordinate system of the novel object are provided as inputs to the network by rendering multiple synthetic views of the object's CAD model. Second, we introduce a novel approach for coarse pose estimation which leverages a network trained to classify whether the pose error between a synthetic rendering and an observed image of the same object can be corrected by the refiner. Third, we introduce a large-scale synthetic dataset of photorealistic images of thousands of objects with diverse visual and shape properties and show that this diversity is crucial to obtain good generalization performance on novel objects. We train our approach on this large synthetic dataset and apply it without retraining to hundreds of novel objects in real images from several pose estimation benchmarks. Our approach achieves state-of-the-art performance on the ModelNet and YCB-Video datasets. An extensive evaluation on the 7 core datasets of the BOP challenge demonstrates that our approach achieves performance competitive with existing approaches that require access to the target objects during training. Code, dataset and trained models are available on the project page: https://megapose6d.github.io/.

Reward learning is a highly active area of research in human-robot interaction (HRI), allowing a broad range of users to specify complex robot behaviour. Experiments with simulated user input play a major role in the development and evaluation of reward learning algorithms due to the availability of a ground truth. In this paper, we review measures for evaluating reward learning algorithms used in HRI, most of which fall into two classes. In a theoretical worst case analysis and several examples, we show that both classes of measures can fail to effectively indicate how good the learned robot behaviour is. Thus, our work contributes to the characterization of sim-to-real gaps of reward learning in HRI.

There is an emerging research interest in addressing the online 3D bin packing problem (3D-BPP), which has a wide range of applications in logistics industry. However, neither heuristic methods nor those based on deep reinforcement learning (DRL) outperform human packers in real logistics scenarios. One important reason is that humans can make corrections after performing inappropriate packing actions by unpacking incorrectly packed items. Inspired by such an unpacking mechanism, we present a DRL-based packing-and-unpacking network (PUN) to learn the synergies between the two actions for the online 3D-BPP. Experimental results demonstrate that PUN achieves the state-of-the-art performance and the supplementary video shows that the system based on PUN can reliably complete the online 3D bin packing task in the real world.

Skill-based reinforcement learning (RL) has emerged as a promising strategy to leverage prior knowledge for accelerated robot learning. Skills are typically extracted from expert demonstrations and are embedded into a latent space from which they can be sampled as actions by a high-level RL agent. However, this \textit{skill space} is expansive, and not all skills are relevant for a given robot state, making exploration difficult. Furthermore, the downstream RL agent is limited to learning structurally similar tasks to those used to construct the skill space. We firstly propose accelerating exploration in the skill space using state-conditioned generative models to directly bias the high-level agent towards only \textit{sampling} skills relevant to a given state based on prior experience. Next, we propose a low-level residual policy for fine-grained \textit{skill adaptation} enabling downstream RL agents to adapt to unseen task variations. Finally, we validate our approach across four challenging manipulation tasks that differ from those used to build the skill space, demonstrating our ability to learn across task variations while significantly accelerating exploration, outperforming prior works.

Monocular visual-inertial odometry (VIO) is a critical problem in robotics and autonomous driving. Traditional methods solve this problem based on filtering or optimization. While being fully interpretable, they rely on manual interference and empirical parameter tuning. On the other hand, learning-based approaches allow for end-to-end training but require a large number of training data to learn millions of parameters. However, the non-interpretable and heavy models hinder the generalization ability. In this paper, we propose a fully differentiable, and interpretable, bird-eye-view (BEV) based VIO model for robots with local planar motion that can be trained without deep neural networks. Specifically, we first adopt Unscented Kalman Filter as a differentiable layer to predict the pitch and roll, where the covariance matrices of noise are learned to filter out the noise of the IMU raw data. Second, the refined pitch and roll are adopted to retrieve a gravity-aligned BEV image of each frame using differentiable camera projection. Finally, a differentiable pose estimator is utilized to estimate the remaining 3 DoF poses between the BEV frames: leading to a 5 DoF pose estimation. Our method allows for learning the covariance matrices end-to-end supervised by the pose estimation loss, demonstrating superior performance to empirical baselines. Experimental results on synthetic and real-world datasets demonstrate that our simple approach is competitive with state-of-the-art methods and generalizes well on unseen scenes.

Robots are well-suited to alleviate the burden of repetitive and tedious manipulation tasks. In many applications though, a robot may be asked to interact with a wide variety of objects, making it hard or even impossible to pre-program visual object classifiers suitable for the task of interest. In this work, we study the problem of learning a classifier for visual objects based on a few examples provided by humans. We frame this problem from the perspective of learning a suitable visual object representation that allows us to distinguish the desired object category from others. Our proposed approach integrates human supervision into the representation learning process by combining contrastive learning with an additional loss function that brings the representations of human examples close to each other in the latent space. Our experiments show that our proposed method performs better than self-supervised and fully supervised learning methods in offline evaluations and can also be used in real-time by a robot in a simplified recycling domain, where recycling streams contain a variety of objects.