CoRL 2021 Papers

Humans have a strong intuitive understanding of the 3D environment around us. The mental model of the physics in our brain applies to objects of different materials and enables us to perform a wide range of manipulation tasks that are far beyond the reach of current robots. In this work, we desire to learn models for dynamic 3D scenes purely from 2D visual observations. Our model combines Neural Radiance Fields (NeRF) and time contrastive learning with an autoencoding framework, which learns viewpoint-invariant 3D-aware scene representations. We show that a dynamics model, constructed over the learned representation space, enables visuomotor control for challenging manipulation tasks involving both rigid bodies and fluids, where the target is specified in a viewpoint different from what the robot operates on. When coupled with an auto-decoding framework, it can even support goal specification from camera viewpoints that are outside the training distribution. We further demonstrate the richness of the learned 3D dynamics model by performing future prediction and novel view synthesis. Finally, we provide detailed ablation studies regarding different system designs and qualitative analysis of the learned representations.

The ability to grasp and manipulate transparent objects is a major challenge for robots. Existing depth cameras have difficulty detecting, localizing, and inferring the geometry of such objects. We propose using neural radiance fields (NeRF) to detect, localize, and infer the geometry of transparent objects with sufficient accuracy to find and grasp them securely. We leverage NeRF's view-independent learned density, place lights to increase specular reflections, and perform a transparency-aware depth-rendering that we feed into the Dex-Net grasp planner. We show how additional lights create specular reflections that improve the quality of the depth map, and test a setup for a robot workcell equipped with an array of cameras to perform transparent object manipulation. We also create synthetic and real datasets of transparent objects in real-world settings, including singulated objects, cluttered tables, and the top rack of a dishwasher. In each setting we show that NeRF and Dex-Net are able to reliably compute robust grasps on transparent objects, achieving 90% and 100% grasp-success rates in physical experiments on an ABB YuMi, on objects where baseline methods fail.

This work develops a learning-based contact estimator for legged robots that bypasses the need for physical sensors and takes multi-modal proprioceptive sensory data as input. Unlike vision-based state estimators, proprioceptive state estimators are agnostic to perceptually degraded situations such as dark or foggy scenes. While some robots are equipped with dedicated physical sensors to detect necessary contact data for state estimation, some robots do not have dedicated contact sensors, and the addition of such sensors is non-trivial without redesigning the hardware. The trained network can estimate contact events on different terrains. The experiments show that a contact-aided invariant extended Kalman filter can generate accurate odometry trajectories compared to a state-of-the-art visual SLAM system, enabling robust proprioceptive odometry.

Long-horizon planning in realistic environments requires the ability to reason over sequential tasks in high-dimensional state spaces with complex dynamics. Classical motion planning algorithms, such as rapidly-exploring random trees, are capable of efficiently exploring large state spaces and computing long-horizon, sequential plans. However, these algorithms are generally challenged with complex, stochastic, and high-dimensional state spaces as well as in the presence of small, topologically complex goal regions, which naturally emerge in tasks that interact with the environment. Machine learning offers a promising solution for its ability to learn general policies that can handle complex interactions and high-dimensional observations. However, these policies are generally limited in horizon length. Our approach, Broadly-Exploring, Local-policy Trees (BELT), merges these two approaches to leverage the strengths of both through a task-conditioned, model-based tree search. BELT uses an RRT-inspired tree search to efficiently explore the state space. Locally, the exploration is guided by a task-conditioned, learned policy capable of performing general short-horizon tasks. This task space can be quite general and abstract; its only requirements are to be sampleable and to well-cover the space of useful tasks. This search is aided by a task-conditioned model that temporally extends dynamics propagation to allow long-horizon search and sequential reasoning over tasks. BELT is demonstrated experimentally to be able to plan long-horizon, sequential trajectories with a goal conditioned policy and generate plans that are robust.

Robotic skills can be learned via imitation learning (IL) using user-provided demonstrations, or via reinforcement learning (RL) using large amounts of autonomously collected experience. Both methods have complementary strengths and weaknesses: RL can reach a high level of performance, but requires exploration, which can be very time consuming and unsafe; IL does not require exploration, but only learns skills that are as good as the provided demonstrations. Can a single method combine the strengths of both approaches? A number of prior methods have aimed to address this question, proposing a variety of techniques that integrate elements of IL and RL. However, scaling up such methods to complex robotic skills that integrate diverse offline data and generalize meaningfully to real-world scenarios still presents a major challenge. In this paper, our aim is to test the scalability of prior IL + RL algorithms and devise a system based on detailed empirical experimentation that combines existing components in the most effective and scalable way. To that end, we present a series of experiments aimed at understanding the implications of each design decision, so as to develop a combined approach that can utilize demonstrations and heterogeneous prior data to attain the best performance on a range of real-world and realistic simulated robotic problems. Our complete method, which we call AW-Opt, combines elements of advantage-weighted regression and QT-Opt, providing a unified approach for integrating demonstrations and offline data for robotic manipulation.

Self-driving vehicles must perceive and predict the future positions of nearby actors to avoid collisions and drive safely. A deep learning module is often responsible for this task, requiring large-scale, high-quality training datasets. Due to high labeling costs, active learning approaches are an appealing solution to maximizing model performance for a given labeling budget. However, despite its appeal, there has been little scientific analysis of active learning approaches for the perception and prediction (P&P) problem. In this work, we study active learning techniques for P&P and find that the traditional active learning formulation is ill-suited. We thus introduce generalizations that ensure that our approach is both cost-aware and allows for fine-grained selection of examples through partially labeled scenes. Extensive experiments on a real-world dataset suggest significant improvements across perception, prediction, and downstream planning tasks.

We introduce a framework for multi-camera 3D object detection. In contrast to existing works, which estimate 3D bounding boxes directly from monocular images or use depth prediction networks to generate input for 3D object detection from 2D information, our method manipulates predictions directly in 3D space. Our architecture extracts 2D features from multiple camera images and then uses a sparse set of 3D object queries to index into these 2D features, linking 3D positions to multi-view images using camera transformation matrices. Finally, our model makes a bounding box prediction per object query, using a set-to-set loss to measure the discrepancy between the ground-truth and the prediction. This top-down approach outperforms its bottom-up counterpart in which object bounding box prediction follows per-pixel depth estimation, since it does not suffer from the compounding error introduced by a depth prediction model. Moreover, our method does not require post-processing such as non-maximum suppression, dramatically improving inference speed. We achieve state-of-the-art performance on the nuScenes autonomous driving benchmark.

Rigid objects make distinctive sounds during manipulation. These sounds are a function of object features, such as shape and material, and of contact forces during manipulation. Being able to infer from sound an object's acoustic properties, how it is being manipulated, and what events it is participating in could augment and complement what robots can perceive from vision, especially in case of occlusion, low visual resolution, poor lighting, or blurred focus. Annotations on sound data are rare. Therefore, existing inference systems mostly include a sound renderer in the loop, and use analysis-by-synthesis to optimize for object acoustic properties. Optimizing parameters with respect to a non-differentiable renderer is slow and hard to scale to complex scenes. We present DiffImpact, a fully differentiable model for sounds rigid objects make during impacts, based on physical principles of impact forces, rigid object vibration, and other acoustic effects. Its differentiability enables gradient-based, efficient joint inference of acoustic properties of the objects and characteristics and timings of each individual impact. DiffImpact can also be plugged in as the decoder of an autoencoder, and trained end-to-end on real audio data, so that the encoder can learn to solve the inverse problem in a self-supervised way. Experiments demonstrate that our model's physics-based inductive biases make it more resource efficient and expressive than state-of-the-art pure learning-based alternatives, on both forward rendering of impact sounds and inverse tasks such as acoustic property inference and blind source separation of impact sounds.

Manipulating articulated objects requires multiple robot arms in general. It is challenging to enable multiple robot arms to collaboratively complete manipulation tasks on articulated objects. In this paper, we present V-MAO, a framework for learning multi-arm manipulation of articulated objects. Our framework includes a variational generative model that learns contact point distribution over object rigid parts for each robot arm. The training signal is obtained from interaction with the simulation environment which is enabled by planning and a novel formulation of object-centric control for articulated objects. We deploy our framework in a customized MuJoCo simulation environment and demonstrate that our framework achieves a high success rate on six different objects and two different robots. We also show that generative modeling can effectively learn the contact point distribution on articulated objects.

In this work, we present and study a training set-up that achieves fast policy generation for real-world robotic tasks by using massive parallelism on a single workstation GPU. We analyze and discuss the impact of different training algorithm components in the massively parallel regime on the final policy performance and training times. In addition, we present a novel game-inspired curriculum that is well suited for training with thousands of simulated robots in parallel. We evaluate the approach by training the quadrupedal robot ANYmal to walk on challenging terrain. The parallel approach allows training policies for flat terrain in under four minutes, and in twenty minutes for uneven terrain. This represents a speedup of multiple orders of magnitude compared to previous work. Finally, we transfer the policies to the real robot to validate the approach.

Autonomous robotic assembly requires a well-orchestrated sequence of high-level actions and smooth manipulation executions. Learning to assemble complex 3D structures remains a challenging problem that requires drawing connections between target designs and building blocks, and creating valid assembly sequences considering structural stability and feasibility. To address the combinatorial complexity of the assembly tasks, we propose a multi-head attention graph representation that can be trained with reinforcement learning (RL) to encode the spatial relations and provide meaningful assembly actions. Combining structured representations with model-free RL and Monte-Carlo planning allows agents to operate with various target shapes and building block types. We design a hierarchical control framework that learns to sequence the building blocks to construct arbitrary 3D designs and ensures their feasibility, as we plan the geometric execution with the robot-in-the-loop. We demonstrate the flexibility of the proposed structured representation and our algorithmic solution in a series of simulated 3D assembly tasks with robotic evaluation, which showcases our method's ability to learn to construct stable structures with a large number of building blocks. Code and videos are available at: https://sites.google.com/view/learn2assemble

Real-world robotic tasks require complex reward functions. When we define the problem the robot needs to solve, we pretend that a designer specifies this complex reward exactly, and it is set in stone from then on. In practice, however, reward design is an iterative process: the designer chooses a reward, eventually encounters an ``````''edge-case'' environment where the reward incentivizes the wrong behavior, revises the reward, and repeats. What would it mean to rethink robotics problems to formally account for this iterative nature of reward design? We propose that the robot not take the specified reward for granted, but rather have uncertainty about it, and account for the future design iterations as future evidence. We contribute an Assisted Reward Design method that speeds up the design process by anticipating and influencing this future evidence: rather than letting the designer eventually encounter failure cases and revise the reward then, the method actively exposes the designer to such environments during the development phase. We test this method in an autonomous driving task and find that it more quickly improves the car's behavior in held-out environments by proposing environments that are ``''edge cases'' for the current reward.

We focus on the problem of developing energy efficient controllers for quadrupedal robots. Animals can actively switch gaits at different speeds to lower their energy consumption. In this paper, we devise a hierarchical learning framework, in which distinctive locomotion gaits and natural gait transitions emerge automatically with a simple reward of energy minimization. We use evolutionary strategies (ES) to train a high-level gait policy that specifies gait patterns of each foot, while the low-level convex MPC controller optimizes the motor commands so that the robot can walk at a desired velocity using that gait pattern. We test our learning framework on a quadruped robot and demonstrate automatic gait transitions, from walking to trotting and to fly-trotting, as the robot increases its speed. We show that the learned hierarchical controller consumes much less energy across a wide range of locomotion speed than baseline controllers.

Today's robots are increasingly interacting with people and need to efficiently learn inexperienced user's preferences. A common framework is to iteratively query the user about which of two presented robot trajectories they prefer. While this minimizes the users effort, a strict choice does not yield any information on how much one trajectory is preferred. We propose scale feedback, where the user utilizes a slider to give more nuanced information. We introduce a probabilistic model on how users would provide feedback and derive a learning framework for the robot. We demonstrate the performance benefit of slider feedback in simulations, and validate our approach in two user studies suggesting that scale feedback enables more effective learning in practice.

We propose a novel approach that biases actions during policy search by lifting the concept of redundancy resolution from multi-DoF robot kinematics to the level of the reward in deep reinforcement learning and evolution strategies. The key idea is to bias the distribution of executed actions in the sense that the immediate reward remains unchanged. The resulting biased actions favor secondary objectives yielding policies that are safer to apply on the real robot. We demonstrate the feasibility of our method, considered as policy search with redundant action bias (PSRAB), in a reaching and a pick-and-lift task with a 7-DoF Franka robot arm trained in RLBench - a recently introduced benchmark for robotic manipulation - using state-of-the-art TD3 deep reinforcement learning and OpenAI's evolutionary strategy. We show that it is a flexible approach without the need of significant fine-tuning and interference with the main objective even across different policy search methods and tasks of different complexity. We evaluate our approach in simulation and on the real robot. Our project website with videos and further results can be found at: https://sites.google.com/view/redundant-action-bias

Human drivers can recognise fast abnormal driving situations to avoid accidents. Similar to humans, automated vehicles are supposed to perform anomaly detection. In this work, we propose the spatio-temporal graph auto-encoder for learning normal driving behaviours. Our innovation is the ability to jointly learn multiple trajectories of a dynamic number of agents. To perform anomaly detection, we first estimate a density function of the learned trajectory feature representation and then detect anomalies in low-density regions. Due to the lack of multi-agent trajectory datasets for anomaly detection in automated driving, we introduce our dataset using a driving simulator for normal and abnormal manoeuvres. Our evaluations show that our approach learns the relation between different agents and delivers promising results compared to the related works. The code, simulation and the dataset are publicly available.

Reinforcement learning is finding its way to real-world problem application, transferring from simulated environments to physical setups. In this work, we implement vision-based alignment of an optical Mach-Zehnder interferometer with a confocal telescope in one arm, which controls the diameter and divergence of the corresponding beam. We use a continuous action space; exponential scaling enables us to handle actions within a range of over two orders of magnitude. Our agent trains only in a simulated environment with domain randomizations. In an experimental evaluation, the agent significantly outperforms an existing solution and a human expert.

We study the problem of learning a range of vision-based manipulation tasks from a large offline dataset of robot interaction. In order to accomplish this, humans need easy and effective ways of specifying tasks to the robot. Goal images are one popular form of task specification, as they are already grounded in the robot's observation space. However, goal images also have a number of drawbacks: they are inconvenient for humans to provide, they can over-specify the desired behavior leading to a sparse reward signal, or under-specify task information in the case of non-goal reaching tasks. Natural language provides a convenient and flexible alternative for task specification, but comes with the challenge of grounding language in the robot's observation space. To scalably learn this grounding we propose to leverage offline pre-collected robotic datasets (including highly sub-optimal, autonomously-collected data) with crowd-sourced natural language labels. With this data, we learn a simple classifier which predicts if a change in state completes a language instruction. This provides a language-conditioned reward function that can then be used for offline multi-task RL. In our experiments, we find that on language-conditioned manipulation tasks our approach outperforms both goal-image specifications and language conditioned imitation techniques by more than 25%, and is able to perform a range of visuomotor tasks from natural language, such as “open the right drawer” and “move the stapler”, on a Franka Emika Panda robot.

A promising approach to solving challenging long-horizon tasks has been to extract behavior priors (skills) by fitting generative models to large offline datasets of demonstrations. However, such generative models inherit the biases of the underlying data and result in poor and unusable skills when trained on imperfect demonstration data. To better align skill extraction with human intent we present Skill Preferences (SkiP), an algorithm that learns a model over human preferences and uses it to extract human-aligned skills from offline data. After extracting human-preferred skills, SkiP also utilizes human feedback to solve downstream tasks with RL. We show that SkiP enables a simulated kitchen robot to solve complex multi-step manipulation tasks and substantially outperforms prior leading RL algorithms with human preferences as well as leading skill extraction algorithms without human preferences.