CoRL 2023 Papers

Large language models (LLMs) exhibit a wide range of promising capabilities --- from step-by-step planning to commonsense reasoning --- that may provide utility for robots, but remain prone to confidently hallucinated predictions. In this work, we present KnowNo, a framework for measuring and aligning the uncertainty of LLM-based planners, such that they know when they don't know, and ask for help when needed. KnowNo builds on the theory of conformal prediction to provide statistical guarantees on task completion while minimizing human help in complex multi-step planning settings. Experiments across a variety of simulated and real robot setups that involve tasks with different modes of ambiguity (for example, from spatial to numeric uncertainties, from human preferences to Winograd schemas) show that KnowNo performs favorably over modern baselines (which may involve ensembles or extensive prompt tuning) in terms of improving efficiency and autonomy, while providing formal assurances. KnowNo can be used with LLMs out-of-the-box without model-finetuning, and suggests a promising lightweight approach to modeling uncertainty that can complement and scale with the growing capabilities of foundation models.

Segmentation and tracking of unseen object instances in discrete frames pose a significant challenge in dynamic industrial robotic contexts, such as distribution warehouses. Here, robots must handle object rearrangements, including shifting, removal, and partial occlusion by new items, and track these items after substantial temporal gaps. The task is further complicated when robots encounter objects beyond their training sets, thereby requiring the ability to segment and track previously unseen items. Considering that continuous observation is often inaccessible in such settings, our task involves working with a discrete set of frames separated by indefinite periods, during which substantial changes to the scene may occur. This task also translates to domestic robotic applications, such as table rearrangement. To address these demanding challenges, we introduce new synthetic and real-world datasets that replicate these industrial and household scenarios. Furthermore, we propose a novel paradigm for joint segmentation and tracking in discrete frames, alongside a transformer module that facilitates efficient inter-frame communication. Our approach significantly outperforms recent methods in our experiments. For additional results and videos, please visit \url{https://sites.google.com/view/stow-corl23}. Code and dataset will be released.

Recent multi-camera 3D object detectors usually leverage temporal information to construct multi-view stereo that alleviates the ill-posed depth estimation. However, they typically assume all the objects are static and directly aggregate features across frames. This work begins with a theoretical and empirical analysis to reveal that ignoring the motion of moving objects can result in serious localization bias. Therefore, we propose to model Dynamic Objects in RecurrenT (DORT) to tackle this problem. In contrast to previous global BirdEye-View (BEV) methods, DORT extracts object-wise local volumes for motion estimation that also alleviates the heavy computational burden. By iteratively refining the estimated object motion and location, the preceding features can be precisely aggregated to the current frame to mitigate the aforementioned adverse effects. The simple framework has two significant appealing properties. It is flexible and practical that can be plugged into most camera-based 3D object detectors. As there are predictions of object motion in the loop, it can easily track objects across frames according to their nearest center distances. Without bells and whistles, DORT outperforms all the previous methods on the nuScenes detection and tracking benchmarks with 62.8% NDS and 57.6% AMOTA, respectively. The source code will be available at https://github.com/OpenRobotLab/DORT.

We present a framework for robot skill acquisition, which 1) efficiently scale up data generation of language-labelled robot data and 2) effectively distills this data down into a robust multi-task language-conditioned visuo-motor policy. For (1), we use a large language model (LLM) to guide high-level planning, and sampling-based robot planners (e.g. motion or grasp samplers) for generating diverse and rich manipulation trajectories. To robustify this data-collection process, the LLM also infers a code-snippet for the success condition of each task, simultaneously enabling the data-collection process to detect failure and retry as well as the automatic labeling of trajectories with success/failure. For (2), we extend the diffusion policy single-task behavior-cloning approach to multi-task settings with language conditioning. Finally, we propose a new multi-task benchmark with 18 tasks across five domains to test long-horizon behavior, common-sense reasoning, tool-use, and intuitive physics. We find that our distilled policy successfully learned the robust retrying behavior in its data collection procedure, while improving absolute success rates by 33.2% on average across five domains. Code, data, and additional qualitative results are available on https://www.cs.columbia.edu/~huy/scalingup/.

We present a self-supervised sensorimotor pre-training approach for robotics. Our model, called RPT, is a Transformer that operates on sequences of sensorimotor tokens. Given a sequence of camera images, proprioceptive robot states, and actions, we encode the sequence into tokens, mask out a subset, and train a model to predict the missing content from the rest. We hypothesize that if a robot can predict the masked-out content it will have acquired a good model of the physical world that can enable it to act. RPT is designed to operate on latent visual representations which makes prediction tractable, enables scaling to larger models, and allows fast inference on a real robot. To evaluate our approach, we collected a dataset of 20,000 real-world trajectories over 9 months using a combination of motion planning and grasping algorithms. We find that sensorimotor pre-training consistently outperforms training from scratch, has favorable scaling properties, and enables transfer across different tasks, environments, and robots.

Reinforcement Learning (RL) methods are typically sample-inefficient, making it challenging to train and deploy RL-policies in real world robots. Even a robust policy trained in simulation requires a real-world deployment to assess their performance. This paper proposes a new approach to evaluate the real-world performance of agent policies prior to deploying them in the real world. Our approach incorporates a simulator along with real-world offline data to evaluate the performance of any policy using the framework of Marginalized Importance Sampling (MIS). Existing MIS methods face two challenges: (1) large density ratios that deviate from a reasonable range and (2) indirect supervision, where the ratio needs to be inferred indirectly, thus exacerbating estimation error. Our approach addresses these challenges by introducing the target policy's occupancy in the simulator as an intermediate variable and learning the density ratio as the product of two terms that can be learned separately. The first term is learned with direct supervision and the second term has a small magnitude, thus making it computationally efficient. We analyze the sample complexity as well as error propagation of our two step-procedure. Furthermore, we empirically evaluate our approach on Sim2Sim environments such as Cartpole, Reacher, and Half-Cheetah. Our results show that our method generalizes well across a variety of Sim2Sim gap, target policies and offline data collection policies. We also demonstrate the performance of our algorithm on a Sim2Real task of validating the performance of a 7 DoF robotic arm using offline data along with the Gazebo simulator.

Training robots with reinforcement learning (RL) typically involves heavy interactions with the environment, and the acquired skills are often sensitive to changes in task environments and robot kinematics. Transfer RL aims to leverage previous knowledge to accelerate learning of new tasks or new body configurations. However, existing methods struggle to generalize to novel robot-task combinations and scale to realistic tasks due to complex architecture design or strong regularization that limits the capacity of the learned policy. We propose Policy Stitching, a novel framework that facilitates robot transfer learning for novel combinations of robots and tasks. Our key idea is to apply modular policy design and align the latent representations between the modular interfaces. Our method allows direct stitching of the robot and task modules trained separately to form a new policy for fast adaptation. Our simulated and real-world experiments on various 3D manipulation tasks demonstrate the superior zero-shot and few-shot transfer learning performances of our method.

Many real-world manipulation tasks consist of a series of subtasks that are significantly different from one another. Such long-horizon, complex tasks highlight the potential of dexterous hands, which possess adaptability and versatility, capable of seamlessly transitioning between different modes of functionality without the need for re-grasping or external tools. However, the challenges arise due to the high-dimensional action space of dexterous hand and complex compositional dynamics of the long-horizon tasks. We present Sequential Dexterity, a general system based on reinforcement learning (RL) that chains multiple dexterous policies for achieving long-horizon task goals. The core of the system is a transition feasibility function that progressively finetunes the sub-policies for enhancing chaining success rate, while also enables autonomous policy-switching for recovery from failures and bypassing redundant stages. Despite being trained only in simulation with a few task objects, our system demonstrates generalization capability to novel object shapes and is able to zero-shot transfer to a real-world robot equipped with a dexterous hand. Code and videos are available at https://sequential-dexterity.github.io.

An outstanding challenge for the widespread deployment of robotic systems like autonomous vehicles is ensuring safe interaction with humans without sacrificing performance. Existing safety methods often neglect the robot’s ability to learn and adapt at runtime, leading to overly conservative behavior. This paper proposes a new closed-loop paradigm for synthesizing safe control policies that explicitly account for the robot’s evolving uncertainty and its ability to quickly respond to future scenarios as they arise, by jointly considering the physical dynamics and the robot’s learning algorithm. We leverage adversarial reinforcement learning for tractable safety analysis under high-dimensional learning dynamics and demonstrate our framework’s ability to work with both Bayesian belief propagation and implicit learning through large pre-trained neural trajectory predictors.

Learning control policies offline from pre-recorded datasets is a promising avenue for solving challenging real-world problems. However, available datasets are typically of mixed quality, with a limited number of the trajectories that we would consider as positive examples; i.e., high-quality demonstrations. Therefore, we propose a novel iterative learning algorithm for identifying expert trajectories in unlabeled mixed-quality robotics datasets given a minimal set of positive examples, surpassing existing algorithms in terms of accuracy. We show that applying behavioral cloning to the resulting filtered dataset outperforms several competitive offline reinforcement learning and imitation learning baselines. We perform experiments on a range of simulated locomotion tasks and on two challenging manipulation tasks on a real robotic system; in these experiments, our method showcases state-of-the-art performance. Our website: \url{https://sites.google.com/view/offline-policy-learning-pubc}.

Differentiable Filters are recursive Bayesian estimators that derive the state transition and measurement models from data alone. Their data-driven nature eschews the need for explicit analytical models, while remaining algorithmic components of the filtering process intact. As a result, the gain mechanism -- a critical component of the filtering process -- remains non-differentiable and cannot be adjusted to the specific nature of the task or context. In this paper, we propose an attention-based Multimodal Differentiable Filter ($\alpha$-MDF) which utilizes modern attention mechanisms to learn multimodal latent representations. Unlike previous differentiable filter frameworks, $\alpha$-MDF substitutes the traditional gain, e.g., the Kalman gain, with a neural attention mechanism. The approach generates specialized, context-dependent gains that can effectively combine multiple input modalities and observed variables. We validate $\alpha$-MDF on a diverse set of robot state estimation tasks in real world and simulation. Our results show $\alpha$-MDF achieves significant reductions in state estimation errors, demonstrating nearly 4-fold improvements compared to state-of-the-art sensor fusion strategies for rigid body robots. Additionally, the $\alpha$-MDF consistently outperforms differentiable filter baselines by up to 45% in soft robotics tasks. The project is available at alpha-mdf.github.io and the codebase is at github.com/ir-lab/alpha-MDF

For 3D object manipulation, methods that build an explicit 3D representation perform better than those relying only on camera images. But using explicit 3D representations like voxels comes at large computing cost, adversely affecting scalability. In this work, we propose RVT, a multi-view transformer for 3D manipulation that is both scalable and accurate. Some key features of RVT are an attention mechanism to aggregate information across views and re-rendering of the camera input from virtual views around the robot workspace. In simulations, we find that a single RVT model works well across 18 RLBench tasks with 249 task variations, achieving $26\%$ higher relative success than the existing state-of-the-art method (PerAct). It also trains 36X faster than PerAct for achieving the same performance and achieves 2.3X the inference speed of PerAct. Further, RVT can perform a variety of manipulation tasks in the real world with just a few ($\sim$10) demonstrations per task. Visual results, code, and trained model are provided at: https://robotic-view-transformer.github.io/.

Our goal is for robots to follow natural language instructions like ``put the towel next to the microwave.'' But getting large amounts of labeled data, i.e. data that contains demonstrations of tasks labeled with the language instruction, is prohibitive. In contrast, obtaining policies that respond to image goals is much easier, because any autonomous trial or demonstration can be labeled in hindsight with its final state as the goal. In this work, we contribute a method that taps into joint image- and goal- conditioned policies with language using only a small amount of language data. Prior work has made progress on this using vision-language models or by jointly training language-goal-conditioned policies, but so far neither method has scaled effectively to real-world robot tasks without significant human annotation. Our method achieves robust performance in the real world by learning an embedding from the labeled data that aligns language not to the goal image, but rather to the desired change between the start and goal images that the instruction corresponds to. We then train a policy on this embedding: the policy benefits from all the unlabeled data, but the aligned embedding provides an *interface* for language to steer the policy. We show instruction following across a variety of manipulation tasks in different scenes, with generalization to language instructions outside of the labeled data.

In this work, we present a scalable reinforcement learning method for training multi-task policies from large offline datasets that can leverage both human demonstrations and autonomously collected data. Our method uses a Transformer to provide a scalable representation for Q-functions trained via offline temporal difference backups. We therefore refer to the method as Q-Transformer. By discretizing each action dimension and representing the Q-value of each action dimension as separate tokens, we can apply effective high-capacity sequence modeling techniques for Q-learning. We present several design decisions that enable good performance with offline RL training, and show that Q-Transformer outperforms prior offline RL algorithms and imitation learning techniques on a large diverse real-world robotic manipulation task suite.

A robot operating in a partially observable environment must perform sensing actions to achieve a goal, such as clearing the objects in front of a shelf to better localize a target object at the back, and estimate its shape for grasping. A POMDP is a principled framework for enabling robots to perform such information-gathering actions. Unfortunately, while robot manipulation domains involve high-dimensional and continuous observation and action spaces, most POMDP solvers are limited to discrete spaces. Recently, POMCPOW has been proposed for continuous POMDPs, which handles continuity using sampling and progressive widening. However, for robot manipulation problems involving camera observations and multiple objects, POMCPOW is too slow to be practical. We take inspiration from the recent work in learning to guide task and motion planning to propose a framework that learns to guide POMCPOW from past planning experience. Our method uses preference learning that utilizes both success and failure trajectories, where the preference label is given by the results of the tree search. We demonstrate the efficacy of our framework in several continuous partially observable robotics domains, including real-world manipulation, where our framework explicitly reasons about the uncertainty in off-the-shelf segmentation and pose estimation algorithms.

General-purpose pre-trained models (``foundation models'') have enabled practitioners to produce generalizable solutions for individual machine learning problems with datasets that are significantly smaller than those required for learning from scratch. Such models are typically trained on large and diverse datasets with weak supervision, consuming much more training data than is available for any individual downstream application. In this paper, we describe the Visual Navigation Transformer (ViNT), a foundation model that aims to bring the success of general-purpose pre-trained models to vision-based robotic navigation. ViNT is trained with a general goal-reaching objective that can be used with any navigation dataset, and employs a flexible Transformer-based architecture to learn navigational affordances and enable efficient adaptation to a variety of downstream navigational tasks. ViNT is trained on a number of existing navigation datasets, comprising hundreds of hours of robotic navigation from a variety of different robotic platforms, and exhibits positive transfer, outperforming specialist models trained on narrower datasets. ViNT can be augmented with diffusion-based goal proposals to explore novel environments, and can solve kilometer-scale navigation problems when equipped with long-range heuristics. ViNT can also be adapted to novel task specifications with a technique inspired by prompt-tuning, where the goal encoder is replaced by an encoding of another task modality (e.g., GPS waypoints or turn-by-turn directions) embedded into the same space of goal tokens. This flexibility and ability to accommodate a variety of downstream problem domains establish ViNT as an effective foundation model for mobile robotics.

3D perceptual representations are well suited for robot manipulation as they easily encode occlusions and simplify spatial reasoning. Many manipulation tasks require high spatial precision in end-effector pose prediction, which typically demands high-resolution 3D feature grids that are computationally expensive to process. As a result, most manipulation policies operate directly in 2D, foregoing 3D inductive biases. In this paper, we introduce Act3D, a manipulation policy transformer that represents the robot’s workspace using a 3D feature field with adaptive resolutions dependent on the task at hand. The model lifts 2D pre-trained features to 3D using sensed depth, and attends to them to compute features for sampled 3D points. It samples 3D point grids in a coarse to fine manner, featurizes them using relative-position attention, and selects where to focus the next round of point sampling. In this way, it efficiently computes 3D action maps of high spatial resolution. Act3D sets a new state-of-the-art in RLBench, an established manipulation benchmark, where it achieves 10% absolute improvement over the previous SOTA 2D multi-view policy on 74 RLBench tasks and 22% absolute improvement with 3x less compute over the previous SOTA 3D policy. We quantify the importance of relative spatial attention, large-scale vision-language pre-trained 2D backbones, and weight tying across coarse-to-fine attentions in ablative experiments.

Robotic manipulation can greatly benefit from the data efficiency, robustness, and predictability of model-based methods if robots can quickly generate models of novel objects they encounter. This is especially difficult when effects like complex joint friction lack clear first-principles models and are usually ignored by physics simulators. Further, numerically-stiff contact dynamics can make common model-building approaches struggle. We propose a method to simultaneously learn contact and continuous dynamics of a novel, possibly multi-link object by observing its motion through contact-rich trajectories. We formulate a system identification process with a loss that infers unmeasured contact forces, penalizing their violation of physical constraints and laws of motion given current model parameters. Our loss is unlike prediction-based losses used in differentiable simulation. Using a new dataset of real articulated object trajectories and an existing cube toss dataset, our method outperforms differentiable simulation and end-to-end alternatives with more data efficiency. See our project page for code, datasets, and media: https://sites.google.com/view/continuous-contact-nets/home