CoRL 2023 Papers

We introduce GROOT, an imitation learning method for learning robust policies with object-centric and 3D priors. GROOT builds policies that generalize beyond their initial training conditions for vision-based manipulation. It constructs object-centric 3D representations that are robust toward background changes and camera views and reason over these representations using a transformer-based policy. Furthermore, we introduce a segmentation correspondence model that allows policies to generalize to new objects at test time. Through comprehensive experiments, we validate the robustness of GROOT policies against perceptual variations in simulated and real-world environments. GROOT’s performance excels in generalization over background changes, camera viewpoint shifts, and the presence of new object instances, whereas both state-of-the-art end-to-end learning methods and object proposal-based approaches fall short. We also extensively evaluate GROOT policies on real robots, where we demonstrate the efficacy under very wild changes in setup. More videos and model details can be found in the appendix and the project website https://ut-austin-rpl.github.io/GROOT.

Simulation parameter settings such as contact models and object geometry approximations are critical to training robust manipulation policies capable of transferring from simulation to real-world deployment. There is often an irreducible gap between simulation and reality: attempting to match the dynamics between simulation and reality may be infeasible and may not lead to policies that perform well in reality for a specific task. We propose AdaptSim, a new task-driven adaptation framework for sim-to-real transfer that aims to optimize task performance in target (real) environments. First, we meta-learn an adaptation policy in simulation using reinforcement learning for adjusting the simulation parameter distribution based on the current policy's performance in a target environment. We then perform iterative real-world adaptation by inferring new simulation parameter distributions for policy training. Our extensive simulation and hardware experiments demonstrate AdaptSim achieving 1-3x asymptotic performance and 2x real data efficiency when adapting to different environments, compared to methods based on Sys-ID and directly training the task policy in target environments.

While there have been significant strides in dexterous manipulation, most of it is limited to benchmark tasks like in-hand reorientation which are of limited utility in the real world. The main benefit of dexterous hands over two-fingered ones is their ability to pickup tools and other objects (including thin ones) and grasp them firmly in order to apply force. However, this task requires both a complex understanding of functional affordances as well as precise low-level control. While prior work obtains affordances from human data this approach doesn't scale to low-level control. Similarly, simulation training cannot give the robot an understanding of real-world semantics. In this paper, we aim to combine the best of both worlds to accomplish functional grasping for in-the-wild objects. We use a modular approach. First, affordances are obtained by matching corresponding regions of different objects and then a low-level policy trained in sim is run to grasp it. We propose a novel application of eigengrasps to reduce the search space of RL using a small amount of human data and find that it leads to more stable and physically realistic motion. We find that eigengrasp action space beats baselines in simulation and outperforms hardcoded grasping in real and matches or outperforms a trained human teleoperator. Videos at https://dexfunc.github.io/.

The ability to detect and analyze failed executions automatically is crucial for an explainable and robust robotic system. Recently, Large Language Models (LLMs) have demonstrated strong reasoning abilities on textual inputs. To leverage the power of LLMs for robot failure explanation, we introduce REFLECT, a framework which queries LLM for failure reasoning based on a hierarchical summary of robot past experiences generated from multisensory observations. The failure explanation can further guide a language-based planner to correct the failure and complete the task. To systematically evaluate the framework, we create the RoboFail dataset with a variety of tasks and failure scenarios. We demonstrate that the LLM-based framework is able to generate informative failure explanations that assist successful correction planning.

Dynamical System (DS) based Learning from Demonstration (LfD) allows learning of reactive motion policies with stability and convergence guarantees from a few trajectories. Yet, current DS learning techniques lack the flexibility to generalize to new task instances as they overlook explicit task parameters that inherently change the underlying demonstrated trajectories. In this work, we propose Elastic-DS, a novel DS learning and generalization approach that embeds task parameters into the Gaussian Mixture Model (GMM) based Linear Parameter Varying (LPV) DS formulation. Central to our approach is the Elastic-GMM, a GMM constrained to SE(3) task-relevant frames. Given a new task instance/context, the Elastic-GMM is transformed with Laplacian Editing and used to re-estimate the LPV-DS policy. Elastic-DS is compositional in nature and can be used to construct flexible multi-step tasks. We showcase its strength on a myriad of simulated and real-robot experiments while preserving desirable control-theoretic guarantees.

Autonomous agriculture applications (e.g., inspection, phenotyping, plucking fruits) require manipulating the plant foliage to look behind the leaves and the branches. Partial visibility, extreme clutter, thin structures, and unknown geometry and dynamics for plants make such manipulation challenging. We tackle these challenges through data-driven methods. We use self-supervision to train SRPNet, a neural network that predicts what space is revealed on execution of a candidate action on a given plant. We use SRPNet with the cross-entropy method to predict actions that are effective at revealing space beneath plant foliage. Furthermore, as SRPNet does not just predict how much space is revealed but also where it is revealed, we can execute a sequence of actions that incrementally reveal more and more space beneath the plant foliage. We experiment with a synthetic (vines) and a real plant (Dracaena) on a physical test-bed across 5 settings including 2 settings that test generalization to novel plant configurations. Our experiments reveal the effectiveness of our overall method, PPG, over a competitive hand-crafted exploration method, and the effectiveness of SRPNet over a hand-crafted dynamics model and relevant ablations. Project website with execution videos, code, data, and models: https://sites.google.com/view/pushingfoliage/.

Human demonstration videos are a widely available data source for robot learning and an intuitive user interface for expressing desired behavior. However, directly extracting reusable robot manipulation skills from unstructured human videos is challenging due to the big embodiment difference and unobserved action parameters. To bridge this embodiment gap, this paper introduces XSkill, an imitation learning framework that 1) discovers a cross-embodiment representation called skill prototypes purely from unlabeled human and robot manipulation videos, 2) transfers the skill representation to robot actions using conditional diffusion policy, and finally, 3) composes the learned skill to accomplish unseen tasks specified by a human prompt video. Our experiments in simulation and real-world environments show that the discovered skill prototypes facilitate both skill transfer and composition for unseen tasks, resulting in a more general and scalable imitation learning framework.

Key to rich, dexterous manipulation in the real world is the ability to coordinate control across two hands. However, while the promise afforded by bimanual robotic systems is immense, constructing control policies for dual arm autonomous systems brings inherent difficulties. One such difficulty is the high-dimensionality of the bimanual action space, which adds complexity to both model-based and data-driven methods. We counteract this challenge by drawing inspiration from humans to propose a novel role assignment framework: a stabilizing arm holds an object in place to simplify the environment while an acting arm executes the task. We instantiate this framework with BimanUal Dexterity from Stabilization (BUDS), which uses a learned restabilizing classifier to alternate between updating a learned stabilization position to keep the environment unchanged, and accomplishing the task with an acting policy learned from demonstrations. We evaluate BUDS on four bimanual tasks of varying complexities on real-world robots, such as zipping jackets and cutting vegetables. Given only 20 demonstrations, BUDS achieves 76.9% task success across our task suite, and generalizes to out-of-distribution objects within a class with a 52.7% success rate. BUDS is 56.0% more successful than an unstructured baseline that instead learns a BC stabilizing policy due to the precision required of these complex tasks. Supplementary material and videos can be found at https://tinyurl.com/stabilizetoact.

Large language models (LLMs) have demonstrated the potential to perform high-level planning. Yet, it remains a challenge for LLMs to comprehend low-level commands, such as joint angle targets or motor torques. This paper proposes an approach to use foot contact patterns as an interface that bridges human commands in natural language and a locomotion controller that outputs these low-level commands. This results in an interactive system for quadrupedal robots that allows the users to craft diverse locomotion behaviors flexibly. We contribute an LLM prompt design, a reward function, and a method to expose the controller to the feasible distribution of contact patterns. The results are a controller capable of achieving diverse locomotion patterns that can be transferred to real robot hardware. Compared with other design choices, the proposed approach enjoys more than 50% success rate in predicting the correct contact patterns and can solve 10 more tasks out of a total of 30 tasks. (\url{https://saytap.github.io})

Despite great strides in language-guided manipulation, existing work has been constrained to table-top settings. Table-tops allow for perfect and consistent camera angles, properties are that do not hold in mobile manipulation. Task plans that involve moving around the environment must be robust to egocentric views and changes in the plane and angle of grasp. A further challenge is ensuring this is all true while still being able to learn skills efficiently from limited data. We propose Spatial-Language Attention Policies (SLAP) as a solution. SLAP uses three-dimensional tokens as the input representation to train a single multi-task, language-conditioned action prediction policy. Our method shows an 80% success rate in the real world across eight tasks with a single model, and a 47.5% success rate when unseen clutter and unseen object configurations are introduced, even with only a handful of examples per task. This represents an improvement of 30% over prior work (20% given unseen distractors and configurations). We see a 4x improvement over baseline in mobile manipulation setting. In addition, we show how SLAPs robustness allows us to execute Task Plans from open-vocabulary instructions using a large language model for multi-step mobile manipulation. For videos, see the website: https://robotslap.github.io

In human-robot collaboration, both human and robotic agents must work together to achieve a set of shared objectives. However, each team member may have individual preferences, or constraints, for how they would like to contribute to the task. Effective teams align their actions to optimize task performance while satisfying each team member's constraints to the greatest extent possible. We propose a framework for representing human and robot contribution constraints in collaborative human-robot tasks. Additionally, we present an approach for learning a human partner's contribution constraint online during a collaborative interaction. We evaluate our approach using a variety of simulated human partners in a collaborative decluttering task. Our results demonstrate that our method improves team performance over baselines with some, but not all, simulated human partners. Furthermore, we conducted a pilot user study to gather preliminary insights into the effectiveness of our approach on task performance and collaborative fluency. Preliminary results suggest that pilot users performed fluently with our method, motivating further investigation into considering preferences that emerge from collaborative interactions.

We present a framework and algorithms to learn controlled dynamics models using neural stochastic differential equations (SDEs)---SDEs whose drift and diffusion terms are both parametrized by neural networks. We construct the drift term to leverage a priori physics knowledge as inductive bias, and we design the diffusion term to represent a distance-aware estimate of the uncertainty in the learned model's predictions---it matches the system's underlying stochasticity when evaluated on states near those from the training dataset, and it predicts highly stochastic dynamics when evaluated on states beyond the training regime. The proposed neural SDEs can be evaluated quickly enough for use in model predictive control algorithms, or they can be used as simulators for model-based reinforcement learning. Furthermore, they make accurate predictions over long time horizons, even when trained on small datasets that cover limited regions of the state space. We demonstrate these capabilities through experiments on simulated robotic systems, as well as by using them to model and control a hexacopter's flight dynamics: A neural SDE trained using only three minutes of manually collected flight data results in a model-based control policy that accurately tracks aggressive trajectories that push the hexacopter's velocity and Euler angles to nearly double the maximum values observed in the training dataset.

With the advent of large language models and large-scale robotic datasets, there has been tremendous progress in high-level decision-making for object manipulation. These generic models are able to interpret complex tasks using language commands, but they often have difficulties generalizing to out-of-distribution objects due to the inability of low-level action primitives. In contrast, existing task-specific models excel in low-level manipulation of unknown objects, but only work for a single type of action. To bridge this gap, we present M2T2, a single model that supplies different types of low-level actions that work robustly on arbitrary objects in cluttered scenes. M2T2 is a transformer model which reasons about contact points and predicts valid gripper poses for different action modes given a raw point cloud of the scene. Trained on a large-scale synthetic dataset with 128K scenes, M2T2 achieves zero-shot sim2real transfer on the real robot, outperforming the baseline system with state-of-the-art task-specific models by about 19% in overall performance and 37.5% in challenging scenes were the object needs to be re-oriented for collision-free placement. M2T2 also achieves state-of-the-art results on a subset of language conditioned tasks in RLBench. Videos of robot experiments on unseen objects in both real world and simulation are available at m2-t2.github.io.

The advancement of robots, particularly those functioning in complex human-centric environments, relies on control solutions that are driven by machine learning. Understanding how learning-based controllers make decisions is crucial since robots are mostly safety-critical systems. This urges a formal and quantitative understanding of the explanatory factors in the interpretability of robot learning. In this paper, we aim to study interpretability of compact neural policies through the lens of disentangled representation. We leverage decision trees to obtain factors of variation [1] for disentanglement in robot learning; these encapsulate skills, behaviors, or strategies toward solving tasks. To assess how well networks uncover the underlying task dynamics, we introduce interpretability metrics that measure disentanglement of learned neural dynamics from a concentration of decisions, mutual information and modularity perspective. We showcase the effectiveness of the connection between interpretability and disentanglement consistently across extensive experimental analysis.

Human drivers can seamlessly adapt their driving decisions across geographical locations with diverse conditions and rules of the road, e.g., left vs. right-hand traffic. In contrast, existing models for autonomous driving have been thus far only deployed within restricted operational domains, \ie, without accounting for varying driving behaviors across locations or model scalability. In this work, we propose GeCo, a single geographically-aware conditional imitation learning (CIL) model that can efficiently learn from heterogeneous and globally distributed data with dynamic environmental, traffic, and social characteristics. Our key insight is to introduce a high-capacity, geo-location-based channel attention mechanism that effectively adapts to local nuances while also flexibly modeling similarities among regions in a data-driven manner. By optimizing a contrastive imitation objective, our proposed approach can efficiently scale across the inherently imbalanced data distributions and location-dependent events. We demonstrate the benefits of our GeCo agent across multiple datasets, cities, and scalable deployment paradigms, \ie, centralized, semi-supervised, and distributed agent training. Specifically, GeCo outperforms CIL baselines by over 14% in open-loop evaluation and 30% in closed-loop testing on CARLA.

We study the problem of offline pre-training and online fine-tuning for reinforcement learning from high-dimensional observations in the context of realistic robot tasks. Recent offline model-free approaches successfully use online fine-tuning to either improve the performance of the agent over the data collection policy or adapt to novel tasks. At the same time, model-based RL algorithms have achieved significant progress in sample efficiency and the complexity of the tasks they can solve, yet remain under-utilized in the fine-tuning setting. In this work, we argue that existing methods for high-dimensional model-based offline RL are not suitable for offline-to-online fine-tuning due to issues with distribution shifts, off-dynamics data, and non-stationary rewards. We propose an on-policy model-based method that can efficiently reuse prior data through model-based value expansion and policy regularization, while preventing model exploitation by controlling epistemic uncertainty. We find that our approach successfully solves tasks from the MetaWorld benchmark, as well as the Franka Kitchen robot manipulation environment completely from images. To our knowledge, MOTO is the first and only method to solve this environment from pixels.

Humans excel in complex long-horizon soft body manipulation tasks via flexible tool use: bread baking requires a knife to slice the dough and a rolling pin to flatten it. Often regarded as a hallmark of human cognition, tool use in autonomous robots remains limited due to challenges in understanding tool-object interactions. Here we develop an intelligent robotic system, RoboCook, which perceives, models, and manipulates elasto-plastic objects with various tools. RoboCook uses point cloud scene representations, models tool-object interactions with Graph Neural Networks (GNNs), and combines tool classification with self-supervised policy learning to devise manipulation plans. We demonstrate that from just 20 minutes of real-world interaction data per tool, a general-purpose robot arm can learn complex long-horizon soft object manipulation tasks, such as making dumplings and alphabet letter cookies. Extensive evaluations show that RoboCook substantially outperforms state-of-the-art approaches, exhibits robustness against severe external disturbances, and demonstrates adaptability to different materials.

Many robotic tasks can have multiple and diverse solutions and, as such, are naturally expressed as goal sets. Examples include navigating to a room, finding a feasible placement location for an object, or opening a drawer enough to reach inside. Using a goal set as a planning objective requires that a model for the objective be explicitly given by the user. However, some goals are intractable to model, leading to uncertainty over the goal (e.g. stable grasping of an object). In this work, we propose a technique for planning directly to a set of sampled goal configurations. We formulate a planning as inference problem with a novel goal likelihood evaluated against the goal samples. To handle the intractable goal likelihood, we employ Generalized Bayesian Inference to approximate the trajectory distribution. The result is a fully differentiable cost which generalizes across a diverse range of goal set objectives for which samples can be obtained. We show that by considering all goal samples throughout the planning process, our method reliably finds plans on manipulation and navigation problems where heuristic approaches fail.

To cope with distribution shifts or non-stationarity in system dynamics, online adaptation algorithms have been introduced to update offline-learned prediction models in real-time. Existing online adaptation methods focus on optimizing the prediction model by utilizing feedback from the latest prediction error. Unfortunately, this feedback-based approach is susceptible to forgetting past information. This work proposes an online adaptation method with feedforward compensation, which uses critical data samples from a memory buffer, instead of the latest samples, to optimize the prediction model. We prove that the proposed approach achieves a smaller error bound compared to previously utilized methods in slow time-varying systems. We conducted experiments on several prediction tasks, which clearly illustrate the superiority of the proposed feedforward adaptation method. Furthermore, our feedforward adaptation technique is capable of estimating an uncertainty bound for predictions.