CoRL 2022 Papers

Bird’s eye view (BEV) semantic segmentation plays a crucial role in spatial sensing for autonomous driving. Although recent literature has made significant progress on BEV map understanding, they are all based on single-agent camera-based systems. These solutions sometimes have difficulty handling occlusions or detecting distant objects in complex traffic scenes. Vehicle-to-Vehicle (V2V) communication technologies have enabled autonomous vehicles to share sensing information, dramatically improving the perception performance and range compared to single-agent systems. In this paper, we propose CoBEVT, the first generic multi-agent multi-camera perception framework that can cooperatively generate BEV map predictions. To efficiently fuse camera features from multi-view and multi-agent data in an underlying Transformer architecture, we design a fused axial attention module (FAX), which captures sparsely local and global spatial interactions across views and agents. The extensive experiments on the V2V perception dataset, OPV2V, demonstrate that CoBEVT achieves state-of-the-art performance for cooperative BEV semantic segmentation. Moreover, CoBEVT is shown to be generalizable to other tasks, including 1) BEV segmentation with single-agent multi-camera and 2) 3D object detection with multi-agent LiDAR systems, achieving state-of-the-art performance with real-time inference speed. The code is available at https://github.com/DerrickXuNu/CoBEVT.

We introduce a neural implicit representation for grasps of objects from multiple robotic hands. Different grasps across multiple robotic hands are encoded into a shared latent space. Each latent vector is learned to decode to the 3D shape of an object and the 3D shape of a robotic hand in a grasping pose in terms of the signed distance functions of the two 3D shapes. In addition, the distance metric in the latent space is learned to preserve the similarity between grasps across different robotic hands, where the similarity of grasps is defined according to contact regions of the robotic hands. This property enables our method to transfer grasps between different grippers including a human hand, and grasp transfer has the potential to share grasping skills between robots and enable robots to learn grasping skills from humans. Furthermore, the encoded signed distance functions of objects and grasps in our implicit representation can be used for 6D object pose estimation with grasping contact optimization from partial point clouds, which enables robotic grasping in the real world.

Decision-making is challenging in robotics environments with continuous object-centric states, continuous actions, long horizons, and sparse feedback. Hierarchical approaches, such as task and motion planning (TAMP), address these challenges by decomposing decision-making into two or more levels of abstraction. In a setting where demonstrations and symbolic predicates are given, prior work has shown how to learn symbolic operators and neural samplers for TAMP with manually designed parameterized policies. Our main contribution is a method for learning parameterized polices in combination with operators and samplers. These components are packaged into modular neuro-symbolic skills and sequenced together with search-then-sample TAMP to solve new tasks. In experiments in four robotics domains, we show that our approach --- bilevel planning with neuro-symbolic skills --- can solve a wide range of tasks with varying initial states, goals, and objects, outperforming six baselines and ablations.

Tactile perception is key for robotics applications such as manipulation. However, tactile data collection is time-consuming, especially when compared to vision. This limits the use of the tactile modality in machine learning solutions in robotics. In this paper, we propose a generative model to simulate realistic tactile sensory data for use in downstream tasks. Starting with easily-obtained camera images, we train Neural Radiance Fields (NeRF) for objects of interest. We then use NeRF-rendered RGB-D images as inputs to a conditional Generative Adversarial Network model (cGAN) to generate tactile images from desired orientations. We evaluate the generated data quantitatively using the Structural Similarity Index and Mean Squared Error metrics, and also using a tactile classification task both in simulation and in the real world. Results show that by augmenting a manually collected dataset, the generated data is able to increase classification accuracy by around 10\%. In addition, we demonstrate that our model is able to transfer from one tactile sensor to another with a small fine-tuning dataset.

Reinforcement learning often suffer from the sparse reward issue in real-world robotics problems. Learning from demonstration (LfD) is an effective way to eliminate this problem, which leverages collected expert data to aid online learning. Prior works often assume that the learning agent and the expert aim to accomplish the same task, which requires collecting new data for every new task. In this paper, we consider the case where the target task is mismatched from but similar with that of the expert. Such setting can be challenging and we found existing LfD methods may encounter a phenomenon called reward signal backward propagation blockages so that the agent cannot be effectively guided by the demonstrations from mismatched task. We propose conservative reward shaping from demonstration (CRSfD), which shapes the sparse rewards using estimated expert value function. To accelerate learning processes, CRSfD guides the agent to conservatively explore around demonstrations. Experimental results of robot manipulation tasks show that our approach outperforms baseline LfD methods when transferring demonstrations collected in a single task to other different but similar tasks.

We study gravitational pivoting, a constrained version of in-hand manipulation, where we aim to control the rotation of an object around the grip point of a parallel gripper. To achieve this, instead of controlling the gripper to avoid slip, we \emph{embrace slip} to allow the object to rotate in-hand. We collect two real-world datasets, a static tracking dataset and a controller-in-the-loop dataset, both annotated with object angle and angular velocity labels. Both datasets contain force-based tactile information on ten different household objects. We train an LSTM model to predict the angular position and velocity of the held object from purely tactile data. We integrate this model with a controller that opens and closes the gripper allowing the object to rotate to desired relative angles. We conduct real-world experiments where the robot is tasked to achieve a relative target angle. We show that our approach outperforms a sliding-window based MLP in a zero-shot generalization setting with unseen objects. Furthermore, we show a 16.6\% improvement in performance when the LSTM model is fine-tuned on a small set of data collected with both the LSTM model and the controller in-the-loop. Code and videos are available at https://rhys-newbury.github.io/projects/pivoting/.

Empowering autonomous agents with 3D understanding for daily objects is a grand challenge in robotics applications. When exploring in an unknown environment, existing methods for object pose estimation are still not satisfactory due to the diversity of object shapes. In this paper, we propose a novel framework for category-level object shape and pose estimation from a single RGB-D image. To handle the intra-category variation, we adopt a semantic primitive representation that encodes diverse shapes into a unified latent space, which is the key to establish reliable correspondences between observed point clouds and estimated shapes. Then, by using a SIM(3)-invariant shape descriptor, we gracefully decouple the shape and pose of an object, thus supporting latent shape optimization of target objects in arbitrary poses. Extensive experiments show that the proposed method achieves SOTA pose estimation performance and better generalization in the real-world dataset. Code and video are available at \url{https://zju3dv.github.io/gCasp}.

Training a reinforcement learning (RL) agent on a real-world robotics task remains generally impractical due to sample inefficiency. Multi-task RL and meta-RL aim to improve sample efficiency by generalizing over a distribution of related tasks. However, doing so is difficult in practice: In multi-task RL, state of the art methods often fail to outperform a degenerate solution that simply learns each task separately. Hypernetworks are a promising path forward since they replicate the separate policies of the degenerate solution while also allowing for generalization across tasks, and are applicable to meta-RL. However, evidence from supervised learning suggests hypernetwork performance is highly sensitive to the initialization. In this paper, we 1) show that hypernetwork initialization is also a critical factor in meta-RL, and that naive initializations yield poor performance; 2) propose a novel hypernetwork initialization scheme that matches or exceeds the performance of a state-of-the-art approach proposed for supervised settings, as well as being simpler and more general; and 3) use this method to show that hypernetworks can improve performance in meta-RL by evaluating on multiple simulated robotics benchmarks.

Humans intuitively solve tasks in versatile ways, varying their behavior in terms of trajectory-based planning and for individual steps. Thus, they can easily generalize and adapt to new and changing environments. Current Imitation Learning algorithms often only consider unimodal expert demonstrations and act in a state-action-based setting, making it difficult for them to imitate human behavior in case of versatile demonstrations. Instead, we combine a mixture of movement primitives with a distribution matching objective to learn versatile behaviors that match the expert’s behavior and versatility. To facilitate generalization to novel task configurations, we do not directly match the agent’s and expert’s trajectory distributions but rather work with concise geometric descriptors which generalize well to unseen task configurations. We empirically validate our method on various robot tasks using versatile human demonstrations and compare to imitation learning algorithms in a state-action setting as well as a trajectory-based setting. We find that the geometric descriptors greatly help in generalizing to new task configurations and that combining them with our distribution-matching objective is crucial for representing and reproducing versatile behavior.

Realistic simulation is key to enabling safe and scalable development of self-driving vehicles. A core component is simulating the sensors so that the entire autonomy system can be tested in simulation. Sensor simulation involves modeling traffic participants, such as vehicles, with high-quality appearance and articulated geometry, and rendering them in real-time. The self-driving industry has employed artists to build these assets. However, this is expensive, slow, and may not reflect reality. Instead, reconstructing assets automatically from sensor data collected in the wild would provide a better path to generating a diverse and large set that provides good real-world coverage. However, current reconstruction approaches struggle on in-the-wild sensor data, due to its sparsity and noise. To tackle these issues, we present CADSim which combines part-aware object-class priors via a small set of CAD models with differentiable rendering to automatically reconstruct vehicle geometry, including articulated wheels, with high-quality appearance. Our experiments show our approach recovers more accurate shape from sparse data compared to existing approaches. Importantly, it also trains and renders efficiently. We demonstrate our reconstructed vehicles in a wide range of applications, including accurate testing of autonomy perception systems.

Commercial and industrial deployments of robot fleets at Amazon, Nimble, Plus One, Waymo, and Zoox query remote human teleoperators when robots are at risk or unable to make task progress. With continual learning, interventions from the remote pool of humans can also be used to improve the robot fleet control policy over time. A central question is how to effectively allocate limited human attention. Prior work addresses this in the single-robot, single-human setting; we formalize the Interactive Fleet Learning (IFL) setting, in which multiple robots interactively query and learn from multiple human supervisors. We propose Return on Human Effort (ROHE) as a new metric and Fleet-DAgger, a family of IFL algorithms. We present an open-source IFL benchmark suite of GPU-accelerated Isaac Gym environments for standardized evaluation and development of IFL algorithms. We compare a novel Fleet-DAgger algorithm to 4 baselines with 100 robots in simulation. We also perform a physical block-pushing experiment with 4 ABB YuMi robot arms and 2 remote humans. Experiments suggest that the allocation of humans to robots significantly affects the performance of the fleet, and that the novel Fleet-DAgger algorithm can achieve up to 8.8x higher ROHE than baselines. See https://tinyurl.com/fleet-dagger for supplemental material.

Hierarchical Imitation Learning (HIL) is an effective way for robots to learn sub-skills from long-horizon unsegmented demonstrations. However, the learned hierarchical structure lacks the mechanism to transfer across multi-tasks or to new tasks, which makes them have to learn from scratch when facing a new situation. Transferring and reorganizing modular sub-skills require fast adaptation ability of the whole hierarchical structure. In this work, we propose Dual Meta Imitation Learning (DMIL), a hierarchical meta imitation learning method where the high-level network and sub-skills are iteratively meta-learned with model-agnostic meta-learning. DMIL uses the likelihood of state-action pairs from each sub-skill as the supervision for the high-level network adaptation and uses the adapted high-level network to determine different data set for each sub-skill adaptation. We theoretically prove the convergence of the iterative training process of DMIL and establish the connection between DMIL and Expectation-Maximization algorithm. Empirically, we achieve state-of-the-art few-shot imitation learning performance on the Meta-world benchmark and competitive results on long-horizon tasks in Kitchen environments.

We present a framework for specifying tasks involving spatial relations between objects using only 5-10 demonstrations and then executing such tasks given point cloud observations of a novel pair of objects in arbitrary initial poses. Our approach structures these rearrangement tasks by assigning a consistent local coordinate frame to the task-relevant object parts, localizing the corresponding coordinate frame on unseen object instances, and executing an action that brings these frames into alignment. We propose an optimization method that uses multiple Neural Descriptor Fields (NDFs) and a single annotated 3D keypoint to assign a set of consistent coordinate frames to the task-relevant object parts. We also propose an energy-based learning scheme to model the joint configuration of the objects that satisfies a desired relational task. We validate our pipeline on three multi-object rearrangement tasks in simulation and on a real robot. Results show that our method can infer relative transformations that satisfy the desired relation between novel objects in unseen initial poses using just a few demonstrations.

We introduce VIOLA, an object-centric imitation learning approach to learning closed-loop visuomotor policies for robot manipulation. Our approach constructs object-centric representations based on general object proposals from a pre-trained vision model. VIOLA uses a transformer-based policy to reason over these representations and attend to the task-relevant visual factors for action prediction. Such object-based structural priors improve deep imitation learning algorithm's robustness against object variations and environmental perturbations. We quantitatively evaluate VIOLA in simulation and on real robots. VIOLA outperforms the state-of-the-art imitation learning methods by $45.8%$ in success rate. It has also been deployed successfully on a physical robot to solve challenging long-horizon tasks, such as dining table arrangement and coffee making. More videos and model details can be found in supplementary material and the project website: https://ut-austin-rpl.github.io/VIOLA/.

We consider planning for mobile robots conducting missions in real-world domains where a priori unknown dynamics affect the robot’s costs and transitions. We study single-episode missions where it is crucial that the robot appropriately trades off exploration and exploitation, such that the learning of the environment dynamics is just enough to effectively complete the mission. Thus, we propose modelling unknown dynamics using Gaussian processes, which provide a principled Bayesian framework for incorporating online observations made by the robot, and using them to predict the dynamics in unexplored areas. We then formulate the problem of mission planning in Markov decision processes under Gaussian process predictions as Bayesian model-based reinforcement learning. This allows us to employ solution techniques that plan more efficiently than previous Gaussian process planning methods are able to. We empirically evaluate the benefits of our formulation in an underwater autonomous vehicle navigation task and robot mission planning in a realistic simulation of a nuclear environment.

As increasing numbers of autonomous vehicles (AVs) are being deployed, it is important to construct a multi-agent self-driving (MASD) system for navigating traffic flows of AVs. In an MASD system, AVs not only navigate themselves to pursue their own goals, but also interact with each other to prevent congestion or collision, especially in scenarios like intersection or lane merging. Multi-agent reinforcement learning (MARL) provides an appealing alternative to generate safe and efficient actions for multiple AVs. However, current MARL methods are limited to describe scenarios where agents interact in either a cooperative of competitive fashion within one episode. Ordinarily, the agents' objectives are defined with a global or team reward function, which fail to deal with the dynamic social preferences (SPs) and mixed motives like human drivers in traffic interactions. To this end, we propose a novel MARL method called Socially-Attentive Policy Optimization (SAPO), which incorporates: (a) a self-attention module to select the most interactive traffic participant for each AV, and (b) a social-aware integration mechanism to integrate objectives of interacting AVs by estimating the dynamic social preferences from their observations. SAPO solves the problem of how to improve the safety and efficiency of MASD systems, by enabling AVs to learn socially-compatible behaviors. Simulation experiments show that SAPO can successfully capture and utilize the variation of the SPs of AVs to achieve superior performance, compared with baselines in MASD scenarios.

In this work, we explore self-supervised visual pre-training on images from diverse, in-the-wild videos for real-world robotic tasks. Like prior work, our visual representations are pre-trained via a masked autoencoder (MAE), frozen, and then passed into a learnable control module. Unlike prior work, we show that the pre-trained representations are effective across a range of real-world robotic tasks and embodiments. We find that our encoder consistently outperforms CLIP (up to 75\%), supervised ImageNet pre-training (up to 81\%), and training from scratch (up to 81\%). Finally, we train a 307M parameter vision transformer on a massive collection of 4.5M images from the Internet and egocentric videos, and demonstrate clearly the benefits of scaling visual pre-training for robot learning.

Deformable object manipulation has great research significance for the robotic community and numerous applications in daily life. In this work, we study how to knot plastic bags that are randomly dropped from the air with a dual-arm robot based on image input. The complex initial configuration and terrible material properties of plastic bags pose challenges to reliable perception and planning. Directly knotting it from random initial states is difficult. To tackle this problem, we propose Iterative Interactive Modeling (IIM) to first adjust the plastic bag to a standing pose with imitation learning to establish a high-confidence keypoint skeleton model, then perform a set of learned motion primitives to knot it. We leverage spatial action maps to accomplish the iterative pick-and-place action and a graph convolutional network to evaluate the adjusted pose during the IIM process. In experiments, we achieve an 85.0% success rate in knotting 4 different plastic bags, including one with no demonstration.

Recently, equivariant neural network models have been shown to improve sample efficiency for tasks in computer vision and reinforcement learning. This paper explores this idea in the context of on-robot policy learning in which a policy must be learned entirely on a physical robotic system without reference to a model, a simulator, or an offline dataset. We focus on applications of Equivariant SAC to robotic manipulation and explore a number of variations of the algorithm. Ultimately, we demonstrate the ability to learn several non-trivial manipulation tasks completely through on-robot experiences in less than an hour or two of wall clock time.