CoRL 2023 Papers

Autonomous racing presents a challenging environment for testing the limits of autonomous vehicle technology. Accurately modeling the vehicle dynamics (with all forces and tires) is critical for high-speed racing, but it remains a difficult task and requires an intricate balance between run-time computational demands and modeling complexity. Researchers have proposed utilizing learning-based methods such as Gaussian Process (GP) for learning vehicle dynamics. However, current approaches often oversimplify the modeling process or apply strong assumptions, leading to unrealistic results that cannot translate to real-world settings. In this paper, we proposed DKL-SKIP method for vehicle dynamics modeling. Our approach outperforms standard GP methods and the N4SID system identification technique in terms of prediction accuracy. In addition to evaluating DKL-SKIP on real-world data, we also evaluate its performance using a high-fidelity autonomous racing AutoVerse simulator. The results highlight the potential of DKL-SKIP as a promising tool for modeling complex vehicle dynamics in both real-world and simulated environments.

Ideally, we would place a robot in a real-world environment and leave it there improving on its own by gathering more experience autonomously. However, algorithms for autonomous robotic learning have been challenging to realize in the real world. While this has often been attributed to the challenge of sample complexity, even sample-efficient techniques are hampered by two major challenges - the difficulty of providing well ``shaped" rewards, and the difficulty of continual reset-free training. In this work, we describe a system for real-world reinforcement learning that enables agents to show continual improvement by training directly in the real world without requiring painstaking effort to hand-design reward functions or reset mechanisms. Our system leverages occasional non-expert human-in-the-loop feedback from remote users to learn informative distance functions to guide exploration while leveraging a simple self-supervised learning algorithm for goal-directed policy learning. We show that in the absence of resets, it is particularly important to account for the current ``reachability" of the exploration policy when deciding which regions of the space to explore. Based on this insight, we instantiate a practical learning system - GEAR, which enables robots to simply be placed in real-world environments and left to train autonomously without interruption. The system streams robot experience to a web interface only requiring occasional asynchronous feedback from remote, crowdsourced, non-expert humans in the form of binary comparative feedback. We evaluate this system on a suite of robotic tasks in simulation and demonstrate its effectiveness at learning behaviors both in simulation and the real world. Project website https://guided-exploration-autonomous-rl.github.io/GEAR/.

Realistic traffic simulation is crucial for developing self-driving software in a safe and scalable manner prior to real-world deployment. Typically, imitation learning (IL) is used to learn human-like traffic agents directly from real-world observations collected offline, but without explicit specification of traffic rules, agents trained from IL alone frequently display unrealistic infractions like collisions and driving off the road. This problem is exacerbated in out-of-distribution and long-tail scenarios. On the other hand, reinforcement learning (RL) can train traffic agents to avoid infractions, but using RL alone results in unhuman-like driving behaviors. We propose Reinforcing Traffic Rules (RTR), a holistic closed-loop learning objective to match expert demonstrations under a traffic compliance constraint, which naturally gives rise to a joint IL + RL approach, obtaining the best of both worlds. Our method learns in closed-loop simulations of both nominal scenarios from real-world datasets as well as procedurally generated long-tail scenarios. Our experiments show that RTR learns more realistic and generalizable traffic simulation policies, achieving significantly better tradeoffs between human-like driving and traffic compliance in both nominal and long-tail scenarios. Moreover, when used as a data generation tool for training prediction models, our learned traffic policy leads to considerably improved downstream prediction metrics compared to baseline traffic agents.

Finding and grasping a target object on a cluttered shelf, especially when the target is occluded by other unknown objects and initially invisible, remains a significant challenge in robotic manipulation. While there have been advances in finding the target object by rearranging surrounding objects using specialized tools, developing algorithms that work with standard robot grippers remains an unresolved issue. In this paper, we introduce a novel framework for finding and grasping the target object using a standard gripper, employing pushing and pick and-place actions. To achieve this, we introduce two indicator functions: (i) an existence function, determining the potential presence of the target, and (ii) a graspability function, assessing the feasibility of grasping the identified target. We then formulate a model-based optimal control problem. The core component of our approach involves leveraging a 3D recognition model, enabling efficient estimation of the proposed indicator functions and their associated dynamics models. Our method succeeds in finding and grasping the target object using a standard robot gripper in both simulations and real-world settings. In particular, we demonstrate the adaptability and robustness of our method in the presence of noise in real-world vision sensor data. The code for our framework is available at https://github.com/seungyeon-k/Search-for-Grasp-public.

Effectively scooping food items poses a substantial challenge for current robotic systems, due to the intricate states and diverse physical properties of food. To address this challenge, we believe in the importance of encoding food items into meaningful representations for effective food scooping. However, the distinctive properties of food items, including deformability, fragility, fluidity, or granularity, pose significant challenges for existing representations. In this paper, we investigate the potential of active perception for learning meaningful food representations in an implicit manner. To this end, we present SCONE, a food-scooping robot learning framework that leverages representations gained from active perception to facilitate food scooping policy learning. SCONE comprises two crucial encoding components: the interactive encoder and the state retrieval module. Through the encoding process, SCONE is capable of capturing properties of food items and vital state characteristics. In our real-world scooping experiments, SCONE excels with a 71% success rate when tasked with 6 previously unseen food items across three different difficulty levels, surpassing state-of-the art methods. This enhanced performance underscores SCONE’s stability, as all food items consistently achieve task success rates exceeding 50%. Additionally, SCONE’s impressive capacity to accommodate diverse initial states enables it to precisely evaluate the present condition of the food, resulting in a compelling scooping success rate. For further information, please visit our website: https://sites.google.com/view/corlscone/home.

Adaptable models could greatly benefit robotic agents operating in the real world, allowing them to deal with novel and varying conditions. While approaches such as Bayesian inference are well-studied frameworks for adapting models to evidence, we build on recent advances in deep generative models which have greatly affected many areas of robotics. Harnessing modern GPU acceleration, we investigate how to quickly adapt the sample generation of neural network models to observations in robotic tasks. We propose a simple and general method that is applicable to various deep generative models and robotic environments. The key idea is to quickly fine-tune the model by fitting it to generated samples matching the observed evidence, using the cross-entropy method. We show that our method can be applied to both autoregressive models and variational autoencoders, and demonstrate its usability in object shape inference from grasping, inverse kinematics calculation, and point cloud completion.

When limited by their own morphologies, humans and some species of animals have the remarkable ability to use objects from the environment toward accomplishing otherwise impossible tasks. Robots might similarly unlock a range of additional capabilities through tool use. Recent techniques for jointly optimizing morphology and control via deep learning are effective at designing locomotion agents. But while outputting a single morphology makes sense for locomotion, manipulation involves a variety of strategies depending on the task goals at hand. A manipulation agent must be capable of rapidly prototyping specialized tools for different goals. Therefore, we propose learning a designer policy, rather than a single design. A designer policy is conditioned on task information and outputs a tool design that helps solve the task. A design-conditioned controller policy can then perform manipulation using these tools. In this work, we take a step towards this goal by introducing a reinforcement learning framework for jointly learning these policies. Through simulated manipulation tasks, we show that this framework is more sample efficient than prior methods in multi-goal or multi-variant settings, can perform zero-shot interpolation or fine-tuning to tackle previously unseen goals, and allows tradeoffs between the complexity of design and control policies under practical constraints. Finally, we deploy our learned policies onto a real robot. Please see our supplementary video and website at https://robotic-tool-design.github.io/ for visualizations.

Sampling-based motion planning algorithms excel at searching global solution paths in geometrically complex settings. However, classical approaches, such as RRT, are difficult to scale beyond low-dimensional search spaces and rely on privileged knowledge e.g. about collision detection and underlying state distances. In this work, we take a step towards the integration of sampling-based planning into the reinforcement learning framework to solve sparse-reward control tasks from high-dimensional inputs. Our method, called VELAP, determines sequences of waypoints through sampling-based exploration in a learned state embedding. Unlike other sampling-based techniques, we iteratively expand a tree-based memory of visited latent areas, which is leveraged to explore a larger portion of the latent space for a given number of search iterations. We demonstrate state-of-the-art results in learning control from offline data in the context of vision-based manipulation under sparse reward feedback. Our method extends the set of available planning tools in model-based reinforcement learning by adding a latent planner that searches globally for feasible paths instead of being bound to a fixed prediction horizon.

Offline reinforcement learning (RL) aims to learn an optimal policy from pre-collected and labeled datasets, which eliminates the time-consuming data collection in online RL. However, offline RL still bears a large burden of specifying/handcrafting extrinsic rewards for each transition in the offline data. As a remedy for the labor-intensive labeling, we propose to endow offline RL tasks with a few expert data and utilize the limited expert data to drive intrinsic rewards, thus eliminating the need for extrinsic rewards. To achieve that, we introduce Calibrated Latent gUidancE (CLUE), which utilizes a conditional variational auto-encoder to learn a latent space such that intrinsic rewards can be directly qualified over the latent space. CLUE's key idea is to align the intrinsic rewards consistent with the expert intention via enforcing the embeddings of expert data to a calibrated contextual representation. We instantiate the expert-driven intrinsic rewards in sparse-reward offline RL tasks, offline imitation learning (IL) tasks, and unsupervised offline RL tasks. Empirically, we find that CLUE can effectively improve the sparse-reward offline RL performance, outperform the state-of-the-art offline IL baselines, and discover diverse skills from static reward-free offline data.

Large language models (LLMs) have demonstrated impressive results in developing generalist planning agents for diverse tasks. However, grounding these plans in expansive, multi-floor, and multi-room environments presents a significant challenge for robotics. We introduce SayPlan, a scalable approach to LLM-based, large-scale task planning for robotics using 3D scene graph (3DSG) representations. To ensure the scalability of our approach, we: (1) exploit the hierarchical nature of 3DSGs to allow LLMs to conduct a "semantic search" for task-relevant subgraphs from a smaller, collapsed representation of the full graph; (2) reduce the planning horizon for the LLM by integrating a classical path planner and (3) introduce an "iterative replanning" pipeline that refines the initial plan using feedback from a scene graph simulator, correcting infeasible actions and avoiding planning failures. We evaluate our approach on two large-scale environments spanning up to 3 floors and 36 rooms with 140 assets and objects and show that our approach is capable of grounding large-scale, long-horizon task plans from abstract, and natural language instruction for a mobile manipulator robot to execute. We provide real robot video demonstrations on our project page https://sayplan.github.io.

Dexterity is often seen as a cornerstone of complex manipulation. Humans are able to perform a host of skills with their hands, from making food to operating tools. In this paper, we investigate these challenges, especially in the case of soft, deformable objects as well as complex, relatively long-horizon tasks. Although, learning such behaviors from scratch can be data inefficient. To circumvent this, we propose a novel approach, DEFT (DExterous Fine-Tuning for Hand Policies), that leverages human-driven priors, which are executed directly in the real world. In order to improve upon these priors, DEFT involves an efficient online optimization procedure. With the integration of human-based learning and online fine-tuning, coupled with a soft robotic hand, DEFT demonstrates success across various tasks, establishing a robust, data-efficient pathway toward general dexterous manipulation. Please see our website at https://dexterousfinetuning.github.io for video results.

In this paper, we study imitation learning under the challenging setting of: (1) only a single demonstration, (2) no further data collection, and (3) no prior task or object knowledge. We show how, with these constraints, imitation learning can be formulated as a combination of trajectory transfer and unseen object pose estimation. To explore this idea, we provide an in-depth study on how state-of-the-art unseen object pose estimators perform for one-shot imitation learning on ten real-world tasks, and we take a deep dive into the effects that camera calibration, pose estimation error, and spatial generalisation have on task success rates. For videos, please visit www.robot-learning.uk/pose-estimation-perspective.

Moving objects to find a fully-occluded target object, known as mechanical search, is a challenging problem in robotics. As objects are often organized semantically, we conjecture that semantic information about object relationships can facilitate mechanical search and reduce search time. Large pretrained vision and language models (VLMs and LLMs) have shown promise in generalizing to uncommon objects and previously unseen real-world environments. In this work, we propose a novel framework called Semantic Mechanical Search (SMS). SMS conducts scene understanding and generates a semantic occupancy distribution explicitly using LLMs. Compared to methods that rely on visual similarities offered by CLIP embeddings, SMS leverages the deep reasoning capabilities of LLMs. Unlike prior work that uses VLMs and LLMs as end-to-end planners, which may not integrate well with specialized geometric planners, SMS can serve as a plug-in semantic module for downstream manipulation or navigation policies. For mechanical search in closed-world settings such as shelves, we compare with a geometric-based planner and show that SMS improves mechanical search performance by 24% across the pharmacy, kitchen, and office domains in simulation and 47.1% in physical experiments. For open-world real environments, SMS can produce better semantic distributions compared to CLIP-based methods, with the potential to be integrated with downstream navigation policies to improve object navigation tasks. Code, data, videos, and Appendix are available here.

While natural language offers a convenient shared interface for humans and robots, enabling robots to interpret and follow language commands remains a longstanding challenge in manipulation. A crucial step to realizing a performant instruction-following robot is achieving semantic manipulation – where a robot interprets language at different specificities, from high-level instructions like "Pick up the stuffed animal" to more detailed inputs like "Grab the left ear of the elephant." To tackle this, we propose Keypoints + Instructions to Execution, a two-step framework for semantic manipulation which attends to both scene semantics (distinguishing between different objects in a visual scene) and object semantics (precisely localizing different parts within an object instance). KITE first grounds an input instruction in a visual scene through 2D image keypoints, providing a highly accurate object-centric bias for downstream action inference. Provided an RGB-D scene observation, KITE then executes a learned keypoint-conditioned skill to carry out the instruction. The combined precision of keypoints and parameterized skills enables fine-grained manipulation with generalization to scene and object variations. Empirically, we demonstrate KITE in 3 real-world environments: long-horizon 6-DoF tabletop manipulation, semantic grasping, and a high-precision coffee-making task. In these settings, KITE achieves a 75%, 70%, and 71% overall success rate for instruction-following, respectively. KITE outperforms frameworks that opt for pre-trained visual language models over keypoint-based grounding, or omit skills in favor of end-to-end visuomotor control, all while being trained from fewer or comparable amounts of demonstrations. Supplementary material, datasets, code, and videos can be found on our website: https://tinyurl.com/kite-site.

Parkour is a grand challenge for legged locomotion that requires robots to overcome various obstacles rapidly in complex environments. Existing methods can generate either diverse but blind locomotion skills or vision-based but specialized skills by using reference animal data or complex rewards. However, \textit{autonomous} parkour requires robots to learn generalizable skills that are both vision-based and diverse to perceive and react to various scenarios. In this work, we propose a system for learning a single end-to-end vision-based parkour policy of diverse parkour skills using a simple reward without any reference motion data. We develop a reinforcement learning method inspired by direct collocation to generate parkour skills, including climbing over high obstacles, leaping over large gaps, crawling beneath low barriers, squeezing through thin slits, and running. We distill these skills into a single vision-based parkour policy and transfer it to a quadrupedal robot using its egocentric depth camera. We demonstrate that our system can empower low-cost quadrupedal robots to autonomously select and execute appropriate parkour skills to traverse challenging environments in the real world. Project website: https://robot-parkour.github.io/

Collaborative perception enables agents to share complementary perceptual information with nearby agents. This can significantly benefit the perception performance and alleviate the issues of single-view perception, such as occlusion and sparsity. Most proposed approaches mainly focus on single modality (especially LiDAR), and not fully exploit the superiority of multi-modal perception. We propose an collaborative perception paradigm, BM2CP, which employs LiDAR and camera to achieve efficient multi-modal perception. BM2CP utilizes LiDAR-guided modal fusion, cooperative depth generation and modality-guided intermediate fusion to acquire deep interactions between modalities and agents. Moreover, it is capable to cope with the special case that one of the sensors is unavailable. Extensive experiments validate that it outperforms the state-of-the-art methods with 50X lower communication volumes in real-world autonomous driving scenarios. Our code is available at supplementary materials.

Learning to produce contact-rich, dynamic behaviors from raw sensory data has been a longstanding challenge in robotics. Prominent approaches primarily focus on using visual and tactile sensing. However, pure vision often fails to capture high-frequency interaction, while current tactile sensors can be too delicate for large-scale data collection. In this work, we propose a data-centric approach to dynamic manipulation that uses an often ignored source of information -- sound. We first collect a dataset of 25k interaction-sound pairs across five dynamic tasks using contact microphones. Then, given this data, we leverage self-supervised learning to accelerate behavior prediction from sound. Our experiments indicate that this self-supervised `pretraining' is crucial to achieving high performance, with a 34.5% lower MSE than plain supervised learning and a 54.3% lower MSE over visual training. Importantly, we find that when asked to generate desired sound profiles, online rollouts of our models on a UR10 robot can produce dynamic behavior that achieves an average of 11.5% improvement over supervised learning on audio similarity metrics. Videos and audio data are best seen on our project website: aurl-anon.github.io

Seamless human-robot manipulation in close proximity relies on accurate forecasts of human motion. While there has been significant progress in learning forecast models at scale, when applied to manipulation tasks, these models accrue high errors at critical transition points leading to degradation in downstream planning performance. Our key insight is that instead of predicting the most likely human motion, it is sufficient to produce forecasts that capture how future human motion would affect the cost of a robot's plan. We present ManiCast, a novel framework that learns cost-aware human forecasts and feeds them to a model predictive control planner to execute collaborative manipulation tasks. Our framework enables fluid, real-time interactions between a human and a 7-DoF robot arm across a number of real-world tasks such as reactive stirring, object handovers, and collaborative table setting. We evaluate both the motion forecasts and the end-to-end forecaster-planner system against a range of learned and heuristic baselines while additionally contributing new datasets. We release our code and datasets at https://portal-cornell.github.io/manicast/.

Proactivity in robot assistance refers to the robot's ability to anticipate user needs and perform assistive actions without explicit requests. This requires understanding user routines, predicting consistent activities, and actively seeking information to predict inconsistent behaviors. We propose SLaTe-PRO (Sequential Latent Temporal model for Predicting Routine Object usage), which improves upon prior state-of-the-art by combining object and user action information, and conditioning object usage predictions on past history. Additionally, we find some human behavior to be inherently stochastic and lacking in contextual cues that the robot can use for proactive assistance. To address such cases, we introduce an interactive query mechanism that can be used to ask queries about the user's intended activities and object use to improve prediction. We evaluate our approach on longitudinal data from three households, spanning 24 activity classes. SLaTe-PRO performance raises the F1 score metric to 0.57 without queries, and 0.60 with user queries, over a score of 0.43 from prior work. We additionally present a case study with a fully autonomous household robot.