CoRL 2023 Papers

A robot deployed in a home over long stretches of time faces a true lifelong learning problem. As it seeks to provide assistance to its users, the robot should leverage any accumulated experience to improve its own knowledge and proficiency. We formalize this setting with a novel formulation of lifelong learning for task and motion planning (TAMP), which endows our learner with the compositionality of TAMP systems. Exploiting the modularity of TAMP, we develop a mixture of generative models that produces candidate continuous parameters for a planner. Whereas most existing lifelong learning approaches determine a priori how data is shared across various models, our approach learns shared and non-shared models and determines which to use online during planning based on auxiliary tasks that serve as a proxy for each model's understanding of a state. Our method exhibits substantial improvements (over time and compared to baselines) in planning success on 2D and BEHAVIOR domains.

4D panoptic segmentation is a challenging but practically useful task that requires every point in a LiDAR point-cloud sequence to be assigned a semantic class label, and individual objects to be segmented and tracked over time. Existing approaches utilize only LiDAR inputs which convey limited information in regions with point sparsity. This problem can, however, be mitigated by utilizing RGB camera images which offer appearance-based information that can reinforce the geometry-based LiDAR features. Motivated by this, we propose 4D-Former: a novel method for 4D panoptic segmentation which leverages both LiDAR and image modalities, and predicts semantic masks as well as temporally consistent object masks for the input point-cloud sequence. We encode semantic classes and objects using a set of concise queries which absorb feature information from both data modalities. Additionally, we propose a learned mechanism to associate object tracks over time which reasons over both appearance and spatial location. We apply 4D-Former to the nuScenes and SemanticKITTI datasets where it achieves state-of-the-art results.

We study how vision-language models trained on Internet-scale data can be incorporated directly into end-to-end robotic control to boost generalization and enable emergent semantic reasoning. Our goal is to enable a single end-to-end trained model to both learn to map robot observations to actions and enjoy the benefits of large-scale pretraining on language and vision-language data from the web. To this end, we propose to co-fine-tune state-of-the-art vision-language models on both robotic trajectory data and Internet-scale vision-language tasks, such as visual question answering. In contrast to other approaches, we propose a simple, general recipe to achieve this goal: in order to fit both natural language responses and robotic actions into the same format, we express the actions as text tokens and incorporate them directly into the training set of the model in the same way as natural language tokens. We refer to such category of models as vision-language-action models (VLA) and instantiate an example of such a model, which we call RT-2. Our extensive evaluation (6k evaluation trials) shows that our approach leads to performant robotic policies and enables RT-2 to obtain a range of emergent capabilities from Internet-scale training. This includes significantly improved generalization to novel objects, the ability to interpret commands not present in the robot training data (such as placing an object onto a particular number or icon), and the ability to perform rudimentary reasoning in response to user commands (such as picking up the smallest or largest object, or the one closest to another object). We further show that incorporating chain of thought reasoning allows RT-2 to perform multi-stage semantic reasoning, for example figuring out which object to pick up for use as an improvised hammer (a rock), or which type of drink is best suited for someone who is tired (an energy drink).

Precise arbitrary trajectory tracking for quadrotors is challenging due to unknown nonlinear dynamics, trajectory infeasibility, and actuation limits. To tackle these challenges, we present DATT, a learning-based approach that can precisely track arbitrary, potentially infeasible trajectories in the presence of large disturbances in the real world. DATT builds on a novel feedforward-feedback-adaptive control structure trained in simulation using reinforcement learning. When deployed on real hardware, DATT is augmented with a disturbance estimator using $\mathcal{L}_1$ adaptive control in closed-loop, without any fine-tuning. DATT significantly outperforms competitive adaptive nonlinear and model predictive controllers for both feasible smooth and infeasible trajectories in unsteady wind fields, including challenging scenarios where baselines completely fail. Moreover, DATT can efficiently run online with an inference time less than 3.2ms, less than 1/4 of the adaptive nonlinear model predictive control baseline.

Seeing-eye robots are very useful tools for guiding visually impaired people, potentially producing a huge societal impact given the low availability and high cost of real guide dogs. Although a few seeing-eye robot systems have already been demonstrated, none considered external tugs from humans, which frequently occur in a real guide dog setting. In this paper, we simultaneously train a locomotion controller that is robust to external tugging forces via Reinforcement Learning~(RL), and an external force estimator via supervised learning. The controller ensures stable walking, and the force estimator enables the robot to respond to the external forces from the human. These forces are used to guide the robot to the global goal, which is unknown to the robot, while the robot guides the human around nearby obstacles via a local planner. Experimental results in simulation and on hardware show that our controller is robust to external forces, and our seeing-eye system can accurately detect force direction. We demonstrate our full seeing-eye robot system on a real quadruped robot with a blindfolded human.

While imitation learning methods have seen a resurgent interest for robotic manipulation, the well-known problem of compounding errors continues to afflict behavioral cloning (BC). Waypoints can help address this problem by reducing the horizon of the learning problem for BC, and thus, the errors compounded over time. However, waypoint labeling is underspecified, and requires additional human supervision. Can we generate waypoints automatically without any additional human supervision? Our key insight is that if a trajectory segment can be approximated by linear motion, the endpoints can be used as waypoints. We propose Automatic Waypoint Extraction (AWE) for imitation learning, a preprocessing module to decompose a demonstration into a minimal set of waypoints which when interpolated linearly can approximate the trajectory up to a specified error threshold. AWE can be combined with any BC algorithm, and we find that AWE can increase the success rate of state-of-the-art algorithms by up to 25% in simulation and by 4-28% on real-world bimanual manipulation tasks, reducing the decision making horizon by up to a factor of 10. Videos and code are available at https://lucys0.github.io/awe/.

Leveraging sensing modalities across diverse spatial and temporal resolutions can improve performance of robotic manipulation tasks. Multi-spatial resolution sensing provides hierarchical information captured at different spatial scales and enables both coarse and precise motions. Simultaneously multi-temporal resolution sensing enables the agent to exhibit high reactivity and real-time control. In this work, we propose a framework for learning generalizable language-conditioned multi-task policies that utilize sensing at different spatial and temporal resolutions using networks of varying capacities to effectively perform real time control of precise and reactive tasks. We leverage off-the-shelf pretrained vision-language models to operate on low-frequency global features along with small non-pretrained models to adapt to high frequency local feedback. Through extensive experiments in 3 domains (coarse, precise and dynamic manipulation tasks), we show that our approach significantly improves ($2\times$ on average) over recent multi-task baselines. Further, our approach generalizes well to visual and geometric variations in target objects and to varying interaction forces.

We propose SCALE, an approach for discovering and learning a diverse set of interpretable robot skills from a limited dataset. Rather than learning a single skill which may fail to capture all the modes in the data, we first identify the different modes via causal reasoning and learn a separate skill for each of them. Our main insight is to associate each mode with a unique set of causally relevant context variables that are discovered by performing causal interventions in simulation. This enables data partitioning based on the causal processes that generated the data, and then compressed skills that ignore the irrelevant variables can be trained. We model each robot skill as a Regional Compressed Option, which extends the options framework by associating a causal process and its relevant variables with the option. Modeled as the skill Data Generating Region, each causal process is local in nature and hence valid over only a subset of the context space. We demonstrate our approach for two representative manipulation tasks: block stacking and peg-in-hole insertion under uncertainty. Our experiments show that our approach yields diverse skills that are compact, robust to domain shifts, and suitable for sim-to-real transfer.

Tracing – estimating the spatial state of – long deformable linear objects such as cables, threads, hoses, or ropes, is useful for a broad range of tasks in homes, retail, factories, construction, transportation, and healthcare. For long deformable linear objects (DLOs or simply cables) with many (over 25) crossings, we present HANDLOOM (Heterogeneous Autoregressive Learned Deformable Linear Object Observation and Manipulation) a learning-based algorithm that fits a trace to a greyscale image of cables. We evaluate HANDLOOM on semi-planar DLO configurations where each crossing involves at most 2 segments. HANDLOOM makes use of neural networks trained with 30,000 simulated examples and 568 real examples to autoregressively estimate traces of cables and classify crossings. Experiments find that in settings with multiple identical cables, HANDLOOM can trace each cable with 80% accuracy. In single-cable images, HANDLOOM can trace and identify knots with 77% accuracy. When HANDLOOM is incorporated into a bimanual robot system, it enables state-based imitation of knot tying with 80% accuracy, and it successfully untangles 64% of cable configurations across 3 levels of difficulty. Additionally, HANDLOOM demonstrates generalization to knot types and materials (rubber, cloth rope) not present in the training dataset with 85% accuracy. Supplementary material, including all code and an annotated dataset of RGB-D images of cables along with ground-truth traces, is at https://sites.google.com/view/cable-tracing.

Some Learning from Demonstrations (LfD) methods handle small mismatches in the action spaces of the teacher and student. Here we address the casewhere the teacher’s morphology is substantially different from that of the student. Our framework, Morphological Adaptation in Imitation Learning (MAIL), bridges this gap allowing us to train an agent from demonstrations by other agents with significantly different morphologies. MAIL learns from suboptimal demonstrations, so long as they provide some guidance towards a desired solution. We demonstrate MAIL on manipulation tasks with rigid and deformable objects including 3D cloth manipulation interacting with rigid obstacles. We train a visual control policy for a robot with one end-effector using demonstrations from a simulated agent with two end-effectors. MAIL shows up to 24% improvement in a normalized performance metric over LfD and non-LfD baselines. It is deployed to a real Franka Panda robot, handles multiple variations in properties for objects (size, rotation, translation), and cloth-specific properties (color, thickness, size, material).

We address the problem of efficiently learning high-level abstractions for task-level robot planning. Existing approaches require large amounts of data and fail to generalize learned abstractions to new environments. To address this, we propose to exploit the independence between spatial and non-spatial state variables in the preconditions of manipulation and navigation skills, mirroring the manipulation-navigation split in robotics research. Given a collection of portable manipulation abstractions (i.e., object-centric manipulation skills paired with matching symbolic representations), we derive an algorithm to automatically generate navigation abstractions that support mobile manipulation planning in a novel environment. We apply our approach to simulated data in AI2Thor and on real robot hardware with a coffee preparation task, efficiently generating plannable representations for mobile manipulators in just a few minutes of robot time, significantly outperforming state-of-the-art baselines.

This paper explores the principles for transforming a quadrupedal robot into a guide robot for individuals with visual impairments. A guide robot has great potential to resolve the limited availability of guide animals that are accessible to only two to three percent of the potential blind or visually impaired (BVI) users. To build a successful guide robot, our paper explores three key topics: (1) formalizing the navigation mechanism of a guide dog and a human, (2) developing a data-driven model of their interaction, and (3) improving user safety. First, we formalize the wayfinding task of the human-guide robot team using Markov Decision Processes based on the literature and interviews. Then we collect real human-robot interaction data from three visually impaired and six sighted people and develop an interaction model called the "Delayed Harness" to effectively simulate the navigation behaviors of the team. Additionally, we introduce an action shielding mechanism to enhance user safety by predicting and filtering out dangerous actions. We evaluate the developed interaction model and the safety mechanism in simulation, which greatly reduce the prediction errors and the number of collisions, respectively. We also demonstrate the integrated system on an AlienGo robot with a rigid harness, by guiding users over 100+ meter trajectories.

We consider a Bayesian approach to offline model-based inverse reinforcement learning (IRL). The proposed framework differs from existing offline model-based IRL approaches by performing simultaneous estimation of the expert's reward function and subjective model of environment dynamics. We make use of a class of prior distributions which parameterizes how accurate the expert’s model of the environment is to develop efficient algorithms to estimate the expert's reward and subjective dynamics in high-dimensional settings. Our analysis reveals a novel insight that the estimated policy exhibits robust performance when the expert is believed (a priori) to have a highly accurate model of the environment. We verify this observation in the MuJoCo environments and show that our algorithms outperform state-of-the-art offline IRL algorithms.

We present ChainedDiffuser, a policy architecture that unifies action keypose prediction and trajectory diffusion generation for learning robot manipulation from demonstrations. Our main innovation is to use a global transformer-based action predictor to predict actions at keyframes, a task that requires multi- modal semantic scene understanding, and to use a local trajectory diffuser to predict trajectory segments that connect predicted macro-actions. ChainedDiffuser sets a new record on established manipulation benchmarks, and outperforms both state-of-the-art keypose (macro-action) prediction models that use motion plan- ners for trajectory prediction, and trajectory diffusion policies that do not predict keyframe macro-actions. We conduct experiments in both simulated and real-world environments and demonstrate ChainedDiffuser’s ability to solve a wide range of manipulation tasks involving interactions with diverse objects.

Imitation learning has been applied to a range of robotic tasks, but can struggle when robots encounter edge cases that are not represented in the training data (i.e., distribution shift). Interactive fleet learning (IFL) mitigates distribution shift by allowing robots to access remote human supervisors during task execution and learn from them over time, but different supervisors may demonstrate the task in different ways. Recent work proposes Implicit Behavior Cloning (IBC), which is able to represent multimodal demonstrations using energy-based models (EBMs). In this work, we propose Implicit Interactive Fleet Learning (IIFL), an algorithm that builds on IBC for interactive imitation learning from multiple heterogeneous human supervisors. A key insight in IIFL is a novel approach for uncertainty quantification in EBMs using Jeffreys divergence. While IIFL is more computationally expensive than explicit methods, results suggest that IIFL achieves a 2.8x higher success rate in simulation experiments and a 4.5x higher return on human effort in a physical block pushing task over (Explicit) IFL, IBC, and other baselines.

Driving safety is a top priority for autonomous vehicles. Orthogonal to prior work handling accident-prone traffic events by algorithm designs at the policy level, we investigate a \textbf{C}losed-loop \textbf{A}dversarial \textbf{T}raining (CAT) framework for safe end-to-end driving in this paper through the lens of environment augmentation. CAT aims to continuously improve the safety of driving agents by training the agent on safety-critical scenarios that are dynamically generated over time. A novel resampling technique is developed to turn log-replay real-world driving scenarios into safety-critical ones via probabilistic factorization, where the adversarial traffic generation is modeled as the multiplication of standard motion prediction sub-problems. Consequently, CAT can launch more efficient physical attacks compared to existing safety-critical scenario generation methods and yields a significantly less computational cost in the iterative learning pipeline. We incorporate CAT into the MetaDrive simulator and validate our approach on hundreds of driving scenarios imported from real-world driving datasets. Experimental results demonstrate that CAT can effectively generate adversarial scenarios countering the agent being trained. After training, the agent can achieve superior driving safety in both log-replay and safety-critical traffic scenarios on the held-out test set. Code and data are available at: https://metadriverse.github.io/cat

We consider the problem of designing distributed collision-avoidance multi-agent control in large-scale environments with potentially moving obstacles, where a large number of agents are required to maintain safety using only local information and reach their goals. This paper addresses the problem of collision avoidance, scalability, and generalizability by introducing graph control barrier functions (GCBFs) for distributed control. The newly introduced GCBF is based on the well-established CBF theory for safety guarantees but utilizes a graph structure for scalable and generalizable decentralized control. We use graph neural networks to learn both neural a GCBF certificate and distributed control. We also extend the framework from handling state-based models to directly taking point clouds from LiDAR for more practical robotics settings. We demonstrated the efficacy of GCBF in a variety of numerical experiments, where the number, density, and traveling distance of agents, as well as the number of unseen and uncontrolled obstacles increase. Empirical results show that GCBF outperforms leading methods such as MAPPO and multi-agent distributed CBF (MDCBF). Trained with only $16$ agents, GCBF can achieve up to $3$ times improvement of success rate (agents reach goals and never encountered in any collisions) on $<500$ agents, and still maintain more than $50\%$ success rates for $>\!1000$ agents when other methods completely fail.

Terrain awareness, i.e., the ability to identify and distinguish different types of terrain, is a critical ability that robots must have to succeed at autonomous off-road navigation. Current approaches that provide robots with this awareness either rely on labeled data which is expensive to collect, engineered features and cost functions that may not generalize, or expert human demonstrations which may not be available. Towards endowing robots with terrain awareness without these limitations, we introduce Self-supervised TErrain Representation LearnING (STERLING), a novel approach for learning terrain representations that relies solely on easy-to-collect, unconstrained (e.g., non-expert), and unlabelled robot experience, with no additional constraints on data collection. STERLING employs a novel multi-modal self-supervision objective through non-contrastive representation learning to learn relevant terrain representations for terrain-aware navigation. Through physical robot experiments in off-road environments, we evaluate STERLING features on the task of preference-aligned visual navigation and find that STERLING features perform on par with fully-supervised approaches and outperform other state-of-the-art methods with respect to preference alignment. Additionally, we perform a large-scale experiment of autonomously hiking a 3-mile long trail which STERLING completes successfully with only two manual interventions, demonstrating its robustness to real-world off-road conditions.

Stowing, the task of placing objects in cluttered shelves or bins, is a common task in warehouse and manufacturing operations. However, this task is still predominantly carried out by human workers as stowing is challenging to automate due to the complex multi-object interactions and long-horizon nature of the task. Previous works typically involve extensive data collection and costly human labeling of semantic priors across diverse object categories. This paper presents a method to learn a generalizable robot stowing policy from predictive model of object interactions and a single demonstration with behavior primitives. We propose a novel framework that utilizes Graph Neural Networks (GNNs) to predict object interactions within the parameter space of behavioral primitives. We further employ primitive-augmented trajectory optimization to search the parameters of a predefined library of heterogeneous behavioral primitives to instantiate the control action. Our framework enables robots to proficiently execute long-horizon stowing tasks with a few keyframes (3-4) from a single demonstration. Despite being solely trained in a simulation, our framework demonstrates remarkable generalization capabilities. It efficiently adapts to a broad spectrum of real-world conditions, including various shelf widths, fluctuating quantities of objects, and objects with diverse attributes such as sizes and shapes.