CoRL 2021 Papers

Reinforcement learning (RL) enables robots to learn skills from interactions with the real world. In practice, the unstructured step-based exploration used in Deep RL -- often very successful in simulation -- leads to jerky motion patterns on real robots. Consequences of the resulting shaky behavior are poor exploration, or even damage to the robot. We address these issues by adapting state-dependent exploration (SDE) to current Deep RL algorithms. To enable this adaptation, we propose two extensions to the original SDE, using more general features and re-sampling the noise periodically, which leads to a new exploration method generalized state-dependent exploration (gSDE). We evaluate gSDE both in simulation, on PyBullet continuous control tasks, and directly on three different real robots: a tendon-driven elastic robot, a quadruped and an RC car. The noise sampling interval of gSDE enables a compromise between performance and smoothness, which allows training directly on the real robots without loss of performance.

To successfully tackle challenging manipulation tasks, autonomous agents must learn a diverse set of skills and how to combine them. Recently, self-supervised agents that set their own abstract goals by exploiting the discovered structure in the environment were shown to perform well on many different tasks. In particular, some of them were applied to learn basic manipulation skills in compositional multi-object environments. However, these methods learn skills without taking the dependencies between objects into account. Thus, the learned skills are difficult to combine in realistic environments. We propose a novel self-supervised agent that estimates relations between environment components and uses them to independently control different parts of the environment state. In addition, the estimated relations between objects can be used to decompose a complex goal into a compatible sequence of subgoals. We show that, by using this framework, an agent can efficiently and automatically learn manipulation tasks in multi-object environments with different relations between objects.

We investigate the visual cross-embodiment imitation setting, in which agents learn policies from videos of other agents (such as humans) demonstrating the same task, but with stark differences in their embodiments -- shape, actions, end-effector dynamics, etc. In this work, we demonstrate that it is possible to automatically discover and learn vision-based reward functions from cross-embodiment demonstration videos that are robust to these differences. Specifically, we present a self-supervised method for Cross-embodiment Inverse Reinforcement Learning (XIRL) that leverages temporal cycle-consistency constraints to learn deep visual embeddings that capture task progression from offline videos of demonstrations across multiple expert agents, each performing the same task differently due to embodiment differences. Prior to our work, producing rewards from self-supervised embeddings typically required alignment with a reference trajectory, which may be difficult to acquire under stark embodiment differences. We show empirically that if the embeddings are aware of task-progress, simply taking the negative distance between the current state and goal state in the learned embedding space is useful as a reward for training policies with reinforcement learning. We find our learned reward function not only works for embodiments seen during training, but also generalizes to entirely new embodiments. Additionally, when transferring real-world human demonstrations to a simulated robot, we find that XIRL is more sample efficient than current best methods.

Today's robotic quadruped systems can robustly walk over a diverse range of rough but continuous terrains, where the terrain elevation varies gradually. Locomotion on discontinuous terrains, such as those with gaps or obstacles, presents a complementary set of challenges. In discontinuous settings, it becomes necessary to plan ahead using visual inputs and to execute agile behaviors beyond robust walking, such as jumps. Such dynamic motion results in significant motion of onboard sensors, which introduces a new set of challenges for real-time visual processing. The requirements of agility and terrain awareness in this setting reinforce the need for robust control. We present Depth-based Impulse Control (DIC), a method for synthesizing highly agile visually-guided locomotion behaviors. DIC affords the flexibility of model-free learning but regularizes behavior through explicit model-based optimization of ground reaction forces. We evaluate performance both in simulation and in the real world.

In this paper, we explore whether a robot can learn to regrasp a diverse set of objects to achieve various desired grasp poses. Regrasping is needed whenever a robot's current grasp pose fails to perform desired manipulation tasks. Endowing robots with such an ability has applications in many domains such as manufacturing or domestic services. Yet, it is a challenging task due to the large diversity of geometry in everyday objects and the high dimensionality of the state and action space. In this paper, we propose a system for robots to take partial point clouds of an object and the supporting environment as inputs and output a sequence of pick-and-place operations to transform an initial object grasp pose to the desired object grasp poses. The key technique includes a neural stable placement predictor and a regrasp graph based solution through leveraging and changing the surrounding environment. We introduce a new and challenging synthetic dataset for learning and evaluating the proposed approach. We demonstrate the effectiveness of our proposed system with both simulator and real-world experiments. More videos and visualization examples are available on our project https://sites.google.com/view/regrasp.

Risk awareness is an important factor to consider when deploying policies on robots in the real-world. Defining the right set of risk metrics can be difficult. In this work, we use a differentiable logic monitor that keeps track of the environmental agents' behaviors and provides a risk metric that the controlled agent can incorporate during planning. We introduce LogicRiskNet, a learning structure that can be constructed from temporal logic formulas describing rules governing a safe agent's behaviors. The network's parameters can be learned from demonstration data. By using temporal logic, the network provides an interpretable architecture that can explain what risk metrics are important to the human. We integrate LogicRiskNet in an inverse optimal control (IOC) framework and show that we can learn to generate trajectory plans that accurately mimic the expert's risk handling behaviors solely from demonstration data. We evaluate our method on a real-world driving dataset.

Modern Reinforcement Learning (RL) algorithms are not sample efficient to train on multi-step tasks in complex domains, impeding their wider deployment in the real world. We address this problem by leveraging the insight that RL models trained to complete one set of tasks can be repurposed to complete related tasks when given just a handful of demonstrations. Based upon this insight, we propose See-SPOT-Run (SSR), a new computational approach to robot learning that enables a robot to complete a variety of real robot tasks in novel problem domains without task-specific training. SSR uses pretrained RL models to create vectors that represent model, task, and action relevance in demonstration and test scenes. SSR then compares these vectors via our Cycle Consistency Distance (CCD) metric to determine the next action to take. SSR completes 58% more task steps and 20% more trials than a baseline few-shot learning method that requires task-specific training. SSR also achieves a four order of magnitude improvement in compute efficiency and a 20% to three order of magnitude improvement in sample efficiency compared to the baseline and to training RL models from scratch. To our knowledge, we are the first to address multi-step tasks from demonstration on a real robot without task-specific training, where both the visual input and action space output are high dimensional. Code will be made available.

Rolling shutter distortion is highly undesirable for photography and computer vision algorithms (e.g., visual SLAM) because pixels can be potentially captured at different times and poses. In this paper, we propose a deep neural network to predict depth and row-wise pose from a single image for rolling shutter correction. Our contribution in this work is to incorporate inertial measurement unit (IMU) data into the pose refinement process, which, compared to the state-of-the-art, greatly enhances the pose prediction. The improved accuracy and robustness make it possible for numerous vision algorithms to use imagery captured by rolling shutter cameras and produce highly accurate results. We also extend a dataset to have real rolling shutter images, IMU data, depth maps, camera poses, and corresponding global shutter images for rolling shutter correction training. We demonstrate the efficacy of the proposed method by evaluating the performance of Direct Sparse Odometry (DSO) algorithm on rolling shutter imagery corrected using the proposed approach. Results show marked improvements of the DSO algorithm over using uncorrected imagery, validating the proposed approach.

Legged locomotion is commonly studied and expressed as a discrete set of gait patterns, like walk, trot, gallop, which are usually treated as given and pre-programmed in legged robots for efficient locomotion at different speeds. However, fixing a set of pre-programmed gaits limits the generality of locomotion. Recent animal motor studies show that these conventional gaits are only prevalent in ideal flat terrain conditions while real-world locomotion is unstructured and more like bouts of intermittent steps. What principles could lead to both structured and unstructured patterns across mammals and how to synthesize them in robots? In this work, we take an analysis-by-synthesis approach and learn to move by minimizing mechanical energy. We demonstrate that learning to minimize energy consumption plays a key role in the emergence of natural locomotion gaits at different speeds in real quadruped robots. The emergent gaits are structured in ideal terrains and look similar to that of horses and sheep. The same approach leads to unstructured gaits in rough terrains which is consistent with the findings in animal motor control. We validate our hypothesis in both simulation and real hardware across natural terrains. Videos at https://energy-locomotion.github.io

Stochastic policy-based deep reinforcement learning (RL) approaches have remarkably succeeded to deal with continuous control tasks. However, applying these methods to manipulation tasks remains a challenge since actuators of a robot manipulator require high dimensional continuous action spaces. In this paper, we propose exploration-bounded exploration actor-critic (EBE-AC), a novel deep RL approach to combine stochastic policy optimization with interpretable human knowledge. The human knowledge is defined as heuristic information based on both physical relationships between a robot and objects and binary signals of whether the robot has achieved certain states. The proposed approach, EBE-AC, combines an off-policy actor-critic algorithm with an entropy maximization based on the heuristic information. On a robotic manipulation task, we demonstrate that EBE-AC outperforms prior state-of-the-art off-policy actor-critic deep RL algorithms in terms of sample efficiency. In addition, we found that EBE-AC can be easily combined with latent information, where EBE-AC with latent information further improved sample efficiency and robustness.

Visual Place Recognition (VPR) for mobile robot global relocalization is a well-studied problem, where contrastive learning based representation training methods have led to state-of-the-art performance. However, these methods are mainly designed for single image based VPR, where sequential information, which is ubiquitous in robotics, is only used as a post-processing step for filtering single image match scores, but is never used to guide the representation learning process itself. In this work, for the first time, we bridge the gap between single image representation learning and sequence matching through "SeqMatchNet" which transforms the single image descriptors such that they become more responsive to the sequence matching metric. We propose a novel triplet loss formulation where the distance metric is based on "sequence matching", that is, the aggregation of temporal order-based Euclidean distances computed using single images. We use the same metric for mining negatives online during the training which helps the optimization process by selecting appropriate positives and harder negatives. To overcome the computational overhead of sequence matching for negative mining, we propose a 2D convolution based formulation of sequence matching for efficiently aggregating distances within a distance matrix computed using single images. We show that our proposed method achieves consistent gains in performance as demonstrated on four benchmark datasets. Source code available at https://github.com/oravus/SeqMatchNet.

Natural language provides an accessible and expressive interface to specify long-term tasks for robotic agents. However, non-experts are likely to specify such tasks with high-level instructions, which abstract over specific robot actions through several layers of abstraction. We propose that key to bridging this gap between language and robot actions over long execution horizons are persistent representations. We propose a persistent spatial semantic representation method, and show how it enables building an agent that performs hierarchical reasoning to effectively execute long-term tasks. We evaluate our approach on the ALFRED benchmark and achieve state-of-the-art results, despite completely avoiding the commonly used step-by-step instructions. https://hlsm-alfred.github.io/

Learning complex manipulation tasks in realistic, obstructed environments is a challenging problem due to hard exploration in the presence of obstacles and high-dimensional visual observations. Prior work tackles the exploration problem by integrating motion planning and reinforcement learning. However, the motion planner augmented policy requires access to state information, which is often not available in the real-world settings. To this end, we propose to distill a state-based motion planner augmented policy to a visual control policy via (1) visual behavioral cloning to remove the motion planner dependency along with its jittery motion, and (2) vision-based reinforcement learning with the guidance of the smoothed trajectories from the behavioral cloning agent. We evaluate our method on three manipulation tasks in obstructed environments and compare it against various reinforcement learning and imitation learning baselines. The results demonstrate that our framework is highly sample-efficient and outperforms the state-of-the-art algorithms. Moreover, coupled with domain randomization, our policy is capable of zero-shot transfer to unseen environment settings with distractors.

Vision is one of the most important perception modalities for legged robots to safely and efficiently navigate uneven terrains, such as stairs and stepping stones. However, training robots to effectively understand high-dimensional visual input for locomotion is a challenging problem. In this work, we propose a framework to train a vision-based locomotion controller which enables a quadrupedal robot to traverse uneven environments. The key idea is to introduce a hierarchical structure with a high-level vision policy and a low-level motion controller. The high-level vision policy takes as inputs the perceived vision signals as well as robot states and outputs the desired footholds and base movement of the robot. These are then realized by the low level motion controller composed of a position controller for swing legs and a MPC-based torque controller for stance legs. We train the vision policy using Deep Reinforcement Learning and demonstrate our approach on a variety of uneven environments such as randomly placed stepping stones, quincuncial piles, stairs, and moving platforms. We also validate our method on a real robot to walk over a series of gaps and climbing up a platform.

The goal of learning from demonstrations is to learn a policy for an agent (imitator) by mimicking the behavior in the demonstrations. Prior works on learning from demonstrations assume that the demonstrations are collected by a demonstrator that has the same dynamics as the imitator. However, in many real-world applications, this assumption is limiting --- to improve the problem of lack of data in robotics, we would like to be able to leverage demonstrations collected from agents with different dynamics. This can be challenging as the demonstrations might not even be feasible for the imitator. Our insight is that we can learn a feasibility metric that captures the likelihood of a demonstration being feasible by the imitator. We develop a feasibility MDP (f-MDP) and derive the feasibility score by learning an optimal policy in the f-MDP. Our proposed feasibility measure encourages the imitator to learn from more informative demonstrations, and disregard the far from feasible demonstrations. Our experiments on four simulated environments and on a real robot show that the policy learned with our approach achieves a higher expected return than prior works. We show the videos of the real robot arm experiments on our website.

We present a method for learning human-robot collaboration policy from human-human collaboration demonstrations. An effective robot assistant must learn to handle diverse human behaviors shown in the demonstrations and be robust when the humans adjust their strategies during online task execution. Our method co-optimizes a human policy and a robot policy in an interactive learning process: the human policy learns to generate diverse and plausible collaborative behaviors from demonstrations while the robot policy learns to assist by estimating the unobserved latent strategy of its human collaborator. Across a 2D strategy game, a human-robot handover task, and a multi-step collaborative manipulation task, our method outperforms the alternatives in both simulated evaluations and when executing the tasks with a real human operator in-the-loop.

We introduce a sample-efficient method for learning state-dependent stiffness control policies for dexterous manipulation. The ability to control stiffness facilitates safe and reliable manipulation by providing compliance and robustness to uncertainties. Most current reinforcement learning approaches to achieve robotic manipulation have exclusively focused on position control, often due to the difficulty of learning high-dimensional stiffness control policies. This difficulty can be partially mitigated via policy guidance such as imitation learning. However, expert stiffness control demonstrations are often expensive or infeasible to record. Therefore, we present an approach to learn Stiffness Control from Augmented Position control Experiences (SCAPE) that bypasses this difficulty by transforming position control demonstrations into approximate, suboptimal stiffness control demonstrations. Then, the suboptimality of the augmented demonstrations is addressed by using complementary techniques that help the agent safely learn from both the demonstrations and reinforcement learning. By using simulation tools and experiments on a robotic testbed, we show that the proposed approach efficiently learns safe manipulation policies and outperforms learned position control policies and several other baseline learning algorithms.

When learning common skills like driving, beginners usually have domain experts standing by to ensure the safety of the learning process. We formulate such learning scheme under the Expert-in-the-loop Reinforcement Learning where a guardian is introduced to safeguard the exploration of the learning agent. While allowing the sufficient exploration in the uncertain environment, the guardian intervenes under dangerous situations and demonstrates the correct actions to avoid potential accidents. Thus ERL enables both exploration and expert's partial demonstration as two training sources. Following such a setting, we develop a novel Expert Guided Policy Optimization (EGPO) method which integrates the guardian in the loop of reinforcement learning. The guardian is composed of an expert policy to generate demonstration and a switch function to decide when to intervene. Particularly, a constrained optimization technique is used to tackle the trivial solution that the agent deliberately behaves dangerously to deceive the expert into taking over. Offline RL technique is further used to learn from the partial demonstration generated by the expert. Safe driving experiments show that our method achieves superior training and test-time safety, outperforms baselines with a substantial margin in sample efficiency, and preserves the generalizabiliy to unseen environments in test-time. Demo video and source code are available at: https://decisionforce.github.io/EGPO/

This paper proposes a correspondence-free method for point cloud rotational registration. We learn an embedding for each point cloud in a feature space that preserves the SO(3)-equivariance property, enabled by recent developments in equivariant neural networks. The proposed shape registration method achieves three major advantages through combining equivariant feature learning with implicit shape models. First, the necessity of data association is removed because of the permutation-invariant property in network architectures similar to PointNet. Second, the registration in feature space can be solved in closed-form using Horn's method due to the SO(3)-equivariance property. Third, the registration is robust to noise in the point cloud because of the joint training of registration and implicit shape reconstruction. The experimental results show superior performance compared with existing correspondence-free deep registration methods.