CoRL 2022 Papers

Robust planning in interactive scenarios requires predicting the uncertain future to make risk-aware decisions. Unfortunately, due to long-tail safety-critical events, the risk is often under-estimated by finite-sampling approximations of probabilistic motion forecasts. This can lead to overconfident and unsafe robot behavior, even with robust planners. Instead of assuming full prediction coverage that robust planners require, we propose to make prediction itself risk-aware. We introduce a new prediction objective to learn a risk-biased distribution over trajectories, so that risk evaluation simplifies to an expected cost estimation under this biased distribution. This reduces sample complexity of the risk estimation during online planning, which is needed for safe real-time performance. Evaluation results in a didactic simulation environment and on a real-world dataset demonstrate the effectiveness of our approach. The code and a demo are available.

In robotics, optimizing controller parameters under safety constraints is an important challenge. Safe Bayesian optimization (BO) quantifies uncertainty in the objective and constraints to safely guide exploration in such settings. Hand-designing a suitable probabilistic model can be challenging however. In the presence of unknown safety constraints, it is crucial to choose reliable model hyper-parameters to avoid safety violations. Here, we propose a data-driven approach to this problem by em meta-learning priors for safe BO from offline data. We build on a meta-learning algorithm, F-PACOH, capable of providing reliable uncertainty quantification in settings of data scarcity. As core contribution, we develop a novel framework for choosing safety-compliant priors in a data-riven manner via empirical uncertainty metrics and a frontier search algorithm. On benchmark functions and a high-precision motion system, we demonstrate that our meta-learnt priors accelerate convergence of safe BO approaches while maintaining safety.

Humans, even at a very early age, can learn visual concepts and understand geometry and layout through active interaction with the environment, and generalize their compositions to complete tasks described by natural languages in novel scenes. To mimic such capability, we propose Embodied Concept Learner (ECL) in an interactive 3D environment. Specifically, a robot agent can ground visual concepts, build semantic maps and plan actions to complete tasks by learning purely from human demonstrations and language instructions, without access to ground-truth semantic and depth supervision from simulations. ECL consists of: (i) an instruction parser that translates the natural languages into executable programs; (ii) an embodied concept learner that grounds visual concepts based on language descriptions; (iii) a map constructor that estimates depth and constructs semantic maps by leveraging the learned concepts; and (iv) a program executor with deterministic policies to execute each program. ECL has several appealing benefits thanks to its modularized design. Firstly, it enables the robotic agent to learn semantics and depth unsupervisedly acting like babies, e.g., ground concepts through active interaction and perceive depth by disparities when moving forward. Secondly, ECL is fully transparent and step-by-step interpretable in long-term planning. Thirdly, ECL could be beneficial for the embodied instruction following (EIF), outperforming previous works on the ALFRED benchmark when the semantic label is not provided. Also, the learned concept can be reused for other downstream tasks, such as reasoning of object states.

Grasp detection based on deep learning has been a research hot spot in recent years. The performance of grasping detection models relies on high-quality, large-scale grasp datasets. Taking comprehensive consideration of quality, extendability, and annotation cost, metric-based simulation methodology is the most promising way to generate grasp annotation. As experts in grasping, human intuitively tends to make grasp decision based both on priori and posteriori knowledge. Inspired by that, a combination of priori and posteriori grasp metrics is intuitively helpful to improve annotation quality. In this paper, we build a hybrid metric group involving both priori and posteriori metrics and propose a grasp evaluator to merge those metrics to approximate human grasp decision capability. Centered on the evaluator, we have constructed an automatic grasp annotation framework, through which a large-scale, high-quality, low annotation cost planar grasp dataset GMD is automatically generated.

A novel framework is proposed to incrementally collect landmark-based graph memory and use the collected memory for image goal navigation. Given a target image to search, an embodied robot utilizes semantic memory to find the target in an unknown environment. In this paper, we present a topological semantic graph memory (TSGM), which consists of (1) a graph builder that takes the observed RGB-D image to construct a topological semantic graph, (2) a cross graph mixer module that takes the collected nodes to get contextual information, and (3) a memory decoder that takes the contextual memory as an input to find an action to the target. On the task of an image goal navigation, TSGM significantly outperforms competitive baselines by +5.0-9.0% on the success rate and +7.0-23.5% on SPL, which means that the TSGM finds efficient paths. Additionally, we demonstrate our method on a mobile robot in real-world image goal scenarios.

A simple gripper can solve more complex manipulation tasks if it can utilize the external environment such as pushing the object against the table or a vertical wall, known as "Extrinsic Dexterity." Previous work in extrinsic dexterity usually has careful assumptions about contacts which impose restrictions on robot design, robot motions, and the variations of the physical parameters. In this work, we develop a system based on reinforcement learning (RL) to address these limitations. We study the task of "Occluded Grasping" which aims to grasp the object in configurations that are initially occluded; the robot needs to move the object into a configuration from which these grasps can be achieved. We present a system with model-free RL that successfully achieves this task using a simple gripper with extrinsic dexterity. The policy learns emergent behaviors of pushing the object against the wall to rotate and then grasp it without additional reward terms on extrinsic dexterity. We discuss important components of the system including the design of the RL problem, multi-grasp training and selection, and policy generalization with automatic curriculum. Most importantly, the policy trained in simulation is zero-shot transferred to a physical robot. It demonstrates dynamic and contact-rich motions with a simple gripper that generalizes across objects with various size, density, surface friction, and shape with a 78% success rate.

Deep reinforcement learning in continuous domains focuses on learning control policies that map states to distributions over actions that ideally concentrate on the optimal choices in each step. In multi-agent navigation problems, the optimal actions depend heavily on the agents' density. Their interaction patterns grow exponentially with respect to such density, making it hard for learning-based methods to generalize. We propose to switch the learning objectives from predicting the optimal actions to predicting sets of admissible actions, which we call control admissibility models (CAMs), such that they can be easily composed and used for online inference for an arbitrary number of agents. We design CAMs using graph neural networks and develop training methods that optimize the CAMs in the standard model-free setting, with the additional benefit of eliminating the need for reward engineering typically required to balance collision avoidance and goal-reaching requirements. We evaluate the proposed approach in multi-agent navigation environments. We show that the CAM models can be trained in environments with only a few agents and be easily composed for deployment in dense environments with hundreds of agents, achieving better performance than state-of-the-art methods.

Learning agile skills is one of the main challenges in robotics. To this end, reinforcement learning approaches have achieved impressive results. These methods require explicit task information in terms of a reward function or an expert that can be queried in simulation to provide a target control output, which limits their applicability. In this work, we propose a generative adversarial method for inferring reward functions from partial and potentially physically incompatible demonstrations for successful skill acquirement where reference or expert demonstrations are not easily accessible. Moreover, we show that by using a Wasserstein GAN formulation and transitions from demonstrations with rough and partial information as input, we are able to extract policies that are robust and capable of imitating demonstrated behaviors. Finally, the obtained skills such as a backflip are tested on an agile quadruped robot called Solo 8 and present faithful replication of hand-held human demonstrations.

Generating safety-critical scenarios, which are crucial yet difficult to collect, provides an effective way to evaluate the robustness of autonomous driving systems. However, the diversity of scenarios and efficiency of generation methods are heavily restricted by the rareness and structure of safety-critical scenarios. Therefore, existing generative models that only estimate distributions from observational data are not satisfying to solve this problem. In this paper, we integrate causality as a prior into the scenario generation and propose a flow-based generative framework, Causal Autoregressive Flow (CausalAF). CausalAF encourages the generative model to uncover and follow the causal relationship among generated objects via novel causal masking operations instead of searching the sample only from observational data. By learning the cause-and-effect mechanism of how the generated scenario causes risk situations rather than just learning correlations from data, CausalAF significantly improves learning efficiency. Extensive experiments on three heterogeneous traffic scenarios illustrate that CausalAF requires much fewer optimization resources to effectively generate safety-critical scenarios. We also show that using generated scenarios as additional training samples empirically improves the robustness of autonomous driving algorithms.

Robots have been used in all sorts of automation, and yet the design of robots remains mainly a manual task. We seek to provide design tools to automate the design of robots themselves. An important challenge in robot design automation is the large and complex design search space which grows exponentially with the number of components, making optimization difficult and sample inefficient. In this work, we present Grammar-guided Latent Space Optimization (GLSO), a framework that transforms design automation into a low-dimensional continuous optimization problem by training a graph variational autoencoder (VAE) to learn a mapping between the graph-structured design space and a continuous latent space. This transformation allows optimization to be conducted in a continuous latent space, where sample efficiency can be significantly boosted by applying algorithms such as Bayesian Optimization. GLSO guides training of the VAE using graph grammar rules and robot world space features, such that the learned latent space focus on valid robots and is easier for the optimization algorithm to explore. Importantly, the trained VAE can be reused to search for designs specialized to multiple different tasks without retraining. We evaluate GLSO by designing robots for a set of locomotion tasks in simulation, and demonstrate that our method outperforms related state-of-the-art robot design automation methods.

Decision Transformers (DT) have drawn upon the success of Transformers by abstracting Reinforcement Learning as a target-return-conditioned, sequence modeling problem. In our work, we claim that the distribution of DT's target-returns represents a series of different tasks that agents must learn to handle. Work in multi-task learning has shown that separating the representations of input data belonging to different tasks can improve performance. We draw from this approach to construct ConDT (Contrastive Decision Transformer). ConDT leverages an enhanced contrastive loss to train a return-dependent transformation of the input embeddings, which we empirically show clusters these embeddings by their return. We find that ConDT significantly outperforms DT in Open-AI Gym domains by 10% and 39% in visually challenging Atari domains.

Trajectory prediction using deep neural networks (DNNs) is an essential component of autonomous driving (AD) systems. However, these methods are vulnerable to adversarial attacks, leading to serious consequences such as collisions. In this work, we identify two key ingredients to defend trajectory prediction models against adversarial attacks including (1) designing effective adversarial training methods and (2) adding domain-specific data augmentation to mitigate the performance degradation on clean data. We demonstrate that our method is able to improve the performance by 46\% on adversarial data and at the cost of only 3\% performance degradation on clean data, compared to the model trained with clean data. Additionally, compared to existing robust methods, our method can improve performance by 21\% on adversarial examples and 9\% on clean data. Our robust model is evaluated with a planner to study its downstream impacts. We demonstrate that our model can significantly reduce the severe accident rates (e.g., collisions and off-road driving).

Reinforcement learning can enable robots to navigate to distant goals while optimizing user-specified reward functions, including preferences for following lanes, staying on paved paths, or avoiding freshly mowed grass. However, online learning from trial-and-error for real-world robots is logistically challenging, and methods that instead can utilize existing datasets of robotic navigation data could be significantly more scalable and enable broader generalization. In this paper, we present ReViND, the first offline RL system for robotic navigation that can leverage previously collected data to optimize user-specified reward functions in the real-world. We evaluate our system for off-road navigation without any additional data collection or fine-tuning, and show that it can navigate to distant goals using only offline training from this dataset, and exhibit behaviors that qualitatively differ based on the user-specified reward function.

A vast literature shows that the learning-based visual perception model is sensitive to adversarial noises, but few works consider the robustness of robotic perception models under widely-existing camera motion perturbations. To this end, we study the robustness of the visual perception model under camera motion perturbations to investigate the influence of camera motion on robotic perception. Specifically, we propose a motion smoothing technique for arbitrary image classification models, whose robustness under camera motion perturbations could be certified. The proposed robustness certification framework based on camera motion smoothing provides effective and scalable robustness guarantees for visual perception modules so that they are applicable to wide robotic applications. As far as we are aware, this is the first work to provide robustness certification for the deep perception module against camera motions, which improves the trustworthiness of robotic perception. A realistic indoor robotic dataset with a dense point cloud map for the entire room, MetaRoom, is introduced for the challenging certifiable robust perception task. We conduct extensive experiments to validate the certification approach via motion smoothing against camera motion perturbations. Our framework guarantees the certified accuracy of 81.7% against camera translation perturbation along depth direction within -0.1m ~ 0.1m. We also validate the effectiveness of our method on the real-world robot by conducting hardware experiments on the robotic arm with an eye-in-hand camera. The code is available at https://github.com/HanjiangHu/camera-motion-smoothing.

We present a novel approach to improve the performance of deep reinforcement learning (DRL) based outdoor robot navigation systems. Most, existing DRL methods are based on carefully designed dense reward functions that learn the efficient behavior in an environment.  We circumvent this issue by working only with sparse rewards (which are easy to design) and propose a novel adaptive Heavy-Tailed Reinforce algorithm for Outdoor Navigation called HTRON. Our main idea is to utilize heavy-tailed policy parametrizations which implicitly induce exploration in sparse reward settings. We evaluate the performance of HTRON against Reinforce, PPO, and TRPO algorithms in three different outdoor scenarios: goal-reaching, obstacle avoidance, and uneven terrain navigation. We observe average an increase of 34.41% in terms of success rate, a 15.15% decrease in the average time steps taken to reach the goal, and a 24.9% decrease in the elevation cost compared to the navigation policies obtained by the other methods. Further, we demonstrate that our algorithm can be transferred directly into a Clearpath Husky robot to perform outdoor terrain navigation in real-world scenarios.

Language-conditioned policies allow robots to interpret and execute human instructions. Learning such policies requires a substantial investment with regards to time and compute resources. Still, the resulting controllers are highly device-specific and cannot easily be transferred to a robot with different morphology, capability, appearance or dynamics. In this paper, we propose a sample-efficient approach for training language-conditioned manipulation policies that allows for rapid transfer across different types of robots. By introducing a novel method, namely Hierarchical Modularity, and adopting supervised attention across multiple sub-modules, we bridge the divide between modular and end-to-end learning and enable the reuse of functional building blocks. In both simulated and real world robot manipulation experiments, we demonstrate that our method outperforms the current state-of-the-art methods and can transfer policies across 4 different robots in a sample-efficient manner. Finally, we show that the functionality of learned sub-modules is maintained beyond the training process and can be used to introspect the robot decision-making process.

Understanding the uncertainty of predictions is a desirable feature for perceptual modules in critical robotic applications. 3D object detectors are neural networks with high-dimensional output space. It suffers from poor calibration in classification and lacks reliable uncertainty estimation in regression. To provide a reliable epistemic uncertainty estimation, we tailor Laplace approximation for 3D object detectors, and propose an Uncertainty Separation and Aggregation pipeline for Bayesian inference. The proposed Laplace-approximation approach can easily convert a deterministic 3D object detector into a Bayesian neural network capable of estimating epistemic uncertainty. The experiment results on the KITTI dataset empirically validate the effectiveness of our proposed methods, and demonstrate that Laplace approximation performs better uncertainty quality than Monte-Carlo Dropout, DeepEnsembles, and deterministic models.

Effective planning of long-horizon deformable object manipulation requires suitable abstractions at both the spatial and temporal levels. Previous methods typically either focus on short-horizon tasks or make strong assumptions that full-state information is available, which prevents their use on deformable objects. In this paper, we propose PlAnning with Spatial-Temporal Abstraction (PASTA), which incorporates both spatial abstraction (reasoning about objects and their relations to each other) and temporal abstraction (reasoning over skills instead of low-level actions). Our framework maps high-dimension 3D observations such as point clouds into a set of latent vectors and plans over skill sequences on top of the latent set representation. We show that our method can effectively perform challenging sequential deformable object manipulation tasks in the real world, which require combining multiple tool-use skills such as cutting with a knife, pushing with a pusher, and spreading dough with a roller. Additional materials can be found at our project website: https://sites.google.com/view/pasta-plan.

Learning tasks only from raw video demonstrations is the current state of the art in robotics visual imitation learning research. The implicit assumption here is that all video demonstrations show an optimal/sub-optimal way of performing the task. What if that is not true? What if one or more videos show a wrong way of executing the task? A task policy learned from such incorrect demonstrations can be potentially unsafe for robots and humans. It is therefore important to analyze the video demonstrations for correctness before handing them over to the policy learning algorithm. This is a challenging task, especially due to the very large state space. This paper proposes a framework to autonomously detect incorrect video demonstrations of sequential tasks consisting of several sub-tasks. We analyze the demonstration pool to identify video(s) for which task-features follow a ‘disruptive’ sequence. We analyze entropy to measure this disruption and – through solving a minmax problem – assign poor weights to incorrect videos. We evaluated the framework with two real-world video datasets: our custom-designed Tea-Making with a YuMi robot and the publicly available 50-Salads. Experimental results show the effectiveness of the proposed framework in detecting incorrect video demonstrations even when they make up 40% of the demonstration set. We also show that various state-of-the-art imitation learning algorithms learn a better policy when incorrect demonstrations are discarded from the training pool.