ICRA 2023 Papers

Dynamic Movement Primitives (DMPs) offer great versatility for encoding, generating and adapting complex end-effector trajectories. DMPs are also very well suited to learning manipulation skills from human demonstration. However, the reactive nature of DMPs restricts their applicability for tool use and object manipulation tasks involving non-holonomic constraints, such as scalpel cutting or catheter steering. In this work, we extend the Cartesian space DMP formulation by adding a coupling term that enforces a pre-defined set of non-holonomic constraints. We obtain the closed-form expression for the constraint forcing term using the Udwadia-Kalaba method. This approach offers a clean and practical solution for guaranteed constraint satisfaction at run-time. Further, the proposed analytical form of the constraint forcing term enables efficient trajectory optimization subject to constraints. We demonstrate the usefulness of this approach by showing how we can learn robotic cutting skills from human demonstration.

The process of designing costmaps for off-road driving tasks is often a challenging and engineering-intensive task. Recent work in costmap design for off-road driving focuses on training deep neural networks to predict costmaps from sensory observations using corpora of expert driving data. However, such approaches are generally subject to over-confident mis-predictions and are rarely evaluated in-the-loop on physical hardware. We present an inverse reinforcement learning-based method of efficiently training deep cost functions that are uncertainty-aware. We do so by leveraging recent advances in highly parallel model-predictive control and robotic risk estimation. In addition to demonstrating improvement at reproducing expert trajectories, we also evaluate the efficacy of these methods in challenging off-road navigation scenarios. We observe that our method significantly outperforms a geometric baseline, resulting in 44% improvement in expert path reconstruction and 57% fewer interventions in practice. We also observe that varying the risk tolerance of the vehicle results in qualitatively different navigation behaviors, especially with respect to higher-risk scenarios such as slopes and tall grass.33More detailed algorithms and additional visualizations are provided in the appendix Appendix (appendix link: tinyurl.com/mtkj63e8)

Observing a human demonstrator manipulate objects provides a rich, scalable and inexpensive source of data for learning robotic policies. However, transferring skills from human videos to a robotic manipulator poses several challenges, not least a difference in action and observation spaces. In this work, we use unlabeled videos of humans solving a wide range of manipulation tasks to learn a task-agnostic reward function for robotic manipulation policies. Thanks to the diversity of this training data, the learned reward function sufficiently generalizes to image observations from a previously unseen robot embodiment and environment to provide a meaningful prior for directed exploration in reinforcement learning. We propose two methods for scoring states relative to a goal image: through direct temporal regression, and through distances in an embedding space obtained with time-contrastive learning. By conditioning the function on a goal image, we are able to reuse one model across a variety of tasks. Unlike prior work on leveraging human videos to teach robots, our method, Human Offline Learned Distances (HOLD) requires neither a priori data from the robot environment, nor a set of task-specific human demonstrations, nor a predefined notion of correspondence across morphologies, yet it is able to accelerate training of several manipulation tasks on a simulated robot arm compared to using only a sparse reward obtained from task completion.

Drivers have a responsibility to exercise reasonable care to avoid collision with other road users. This assumed responsibility allows interacting agents to maintain safety without explicit coordination. Thus to enable safe autonomous vehicle (AV) interactions, AVs must understand what their responsibilities are to maintain safety and how they affect the safety of nearby agents. In this work we seek to understand how responsibility is shared in multi-agent settings where an autonomous agent is interacting with human counterparts. We introduce Responsibility-Aware Control Barrier Functions (RA-CBFs) and present a method to learn responsibility allocations from data. By combining safety-critical control and learning- based techniques, RA-CBFs allow us to account for scene- dependent responsibility allocations and synthesize safe and efficient driving behaviors without making worst-case assumptions that typically result in overly-conservative behaviors. We test our framework using real-world driving data and demonstrate its efficacy as a tool for both safe control and forensic analysis of unsafe driving.

Flat objects with negligible thicknesses like books and disks are challenging to be grasped by the robot because of the width limit of the robot's gripper, especially when they are in cluttered environments. Pre-grasp manipulation is conducive to rearranging objects on the table and moving the flat objects to the table edge, making them graspable. In this paper, we formulate this task as Parameterized Action Markov Decision Process, and a novel method based on deep reinforcement learning is proposed to address this problem by introducing sliding primitives as actions. A weight-sharing policy network is utilized to predict the sliding primitive's parameters for each object, and a Q-network is adopted to select the acted object among all the candidates on the table. Meanwhile, via integrating a curriculum learning scheme, our method can be scaled to cluttered environments with more objects. In both simulation and real-world experiments, our method surpasses the existing methods and achieves pre-grasp manipulation with higher task success rates and fewer action steps. Without fine-tuning, it can be generalized to novel shapes and household objects with more than 85% success rates in the real world. Videos and supplementary materials are available at https://sites.google.com/view/pre-grasp-sliding.

Physically assistive robots and exoskeletons have great potential to help humans with a wide variety of collaborative tasks. However, a challenging aspect of the control of such devices is to accurately model or predict human behaviour, which can be highly individual and personalised. In this work, we implement a framework for learning subject-specific models of underlying human motion strategies using inverse musculoskeletal optimal control. We apply this framework to a specific motion task: the sit-to-stand transition. By collecting sit-to-stand data from 4 subjects with and without perturbations, we show that humans modulate their sit-to-stand strategy in the presence of instability, and learn the corresponding models of these strategies. In the future, the personalised motion strategies resulting from this framework could be used to inform the design of real-time assistance strategies for human-robot collaboration problems.

While current systems for autonomous robot navigation can produce safe and efficient motion plans in static environments, they usually generate suboptimal behaviors when multiple robots must navigate together in confined spaces. For example, when two robots meet each other in a narrow hallway, they may either turn around to find an alternative route or collide with each other. This paper presents a new approach to navigation that allows two robots to pass each other in a narrow hallway without colliding, stopping, or waiting. Our approach, Perceptual Hallucination for Hallway Passing (PHHP), learns to synthetically generate virtual obstacles (i.e., perceptual hallucination) to facilitate passing in narrow hallways by multiple robots that utilize otherwise standard autonomous navigation systems. Our experiments on physical robots in a variety of hallways show improved performance compared to multiple baselines.

Recently, neural control policies have outperformed existing model-based planning-and-control methods for autonomously navigating quadrotors through cluttered environments in minimum time. However, they are not perception aware, a crucial requirement in vision-based navigation due to the camera's limited field of view and the underactuated nature of a quadrotor. We propose a learning-based system that achieves perception-aware, agile flight in cluttered environments. Our method combines imitation learning with reinforcement learning (RL) by leveraging a privileged learning-by-cheating framework. Using RL, we first train a perception-aware teacher policy with full-state information to fly in minimum time through cluttered environments. Then, we use imitation learning to distill its knowledge into a vision-based student policy that only perceives the environment via a camera. Our approach tightly couples perception and control, showing a significant advantage in computation speed (10×faster) and success rate. We demonstrate the closed-loop control performance using hardware-in-the-loop simulation. Video: https://youtu.be/9q059CFGcVA

Given a natural language instruction and an input scene, our goal is to train a model to output a manipulation program that can be executed by the robot. Prior approaches for this task possess one of the following limitations: (i) rely on hand-coded symbols for concepts limiting generalization beyond those seen during training [1] (ii) infer action sequences from instructions but require dense sub-goal supervision [2] or (iii) lack semantics required for deeper object-centric reasoning inherent in interpreting complex instructions [3]. In contrast, our approach can handle linguistic as well as perceptual variations, end-to-end trainable and requires no intermediate supervision. The proposed model uses symbolic reasoning constructs that operate on a latent neural object-centric representation, allowing for deeper reasoning over the input scene. Central to our approach is a modular structure consisting of a hierarchical instruction parser and an action simulator to learn disentangled action representations. Our experiments on a simulated environment with a 7-DOF manipulator, consisting of instructions with varying number of steps and scenes with different number of objects, demonstrate that our model is robust to such variations and significantly outperforms baselines, particularly in the generalization settings. The code, dataset and experiment videos are available at https://nsrmp.github.io

Control policy learning for modular robot locomotion has previously been limited to proprioceptive feedback and flat terrain. This paper develops policies for modular systems with vision traversing more challenging environments. These modular robots can be reconfigured to form many different designs, where each design needs a controller to function. Though one could create a policy for individual designs and environments, such an approach is not scalable given the wide range of potential designs and environments. To address this challenge, we create a visual-motor policy that can generalize to both new designs and environments. The policy itself is modular, in that it is divided into components, each of which corresponds to a type of module (e.g., a leg, wheel, or body). The policy components can be recombined during training to learn to control multiple designs. We develop a deep reinforcement learning algorithm where visual observations are input to a modular policy interacting with multiple environments at once. We apply this algorithm to train robots with combinations of legs and wheels, then demonstrate the policy controlling real robots climbing stairs and curbs.

Robotic locomotion is often approached with the goal of maximizing robustness and reactivity by increasing motion control frequency. We challenge this intuitive notion by demonstrating robust and dynamic locomotion with a learned motion controller executing at as low as 8 Hz on a real ANYmal C quadruped. The robot is able to robustly and repeatably achieve a high heading velocity of 1.5 ms-1, traverse uneven terrain, and resist unexpected external perturbations. We further present a comparative analysis of deep reinforcement learning (RL) based motion control policies trained and executed at frequencies ranging from 5 Hz to 200 Hz. We show that low-frequency policies are less sensitive to actuation latencies and variations in system dynamics. This is to the extent that a successful sim- to-real transfer can be performed even without any dynamics randomization or actuation modeling. We support this claim through a set of rigorous empirical evaluations. Moreover, to assist reproducibility, we provide the training and deployment code along with an extended analysis at https://ori-drs.github.io/lfmc/.

We address the problem of grasping unknown objects identified from top-down images with a parallel gripper. When no object 3D model is available, the state-of-the-art grasp generators identify the best candidate locations for planar grasps using the RGBD image. However, while they generate the Cartesian location and orientation of the gripper, the height of the grasp center is often determined by heuristics based on the highest point in the depth map, which leads to unsuccessful grasps when the objects are not thick, or have transparencies or curved shapes. In this paper, we propose to learn a regressor that predicts the best grasp height based from the image. We train this regressor with a dataset that is automatically acquired thanks to the DIGIT optical tactile sensors, which can evaluate grasp success and stability. Using our predictor, the grasping success is improved by 6% for all objects, by 16% on average on difficult objects, and by 40% for objects that are notably very difficult to grasp (e.g., transparent, curved, thin).

The skill of pivoting an object with a robotic system is challenging for the external forces that act on the system, mainly given by contact interaction. The complexity increases when the same skills are required to generalize across different objects. This paper proposes a framework for learning robust and generalizable pivoting skills, which consists of three steps. First, we learn a pivoting policy on an “unitary” object using Reinforcement Learning (RL). Then, we obtain the object's feature space by supervised learning to encode the kinematic properties of arbitrary objects. Finally, to adapt the unitary policy to multiple objects, we learn data-driven projections based on the object features to adjust the state and action space of the new pivoting task. The proposed approach is entirely trained in simulation. It requires only one depth image of the object and can zero-shot transfer to real-world objects. We demonstrate robustness to sim-to-real transfer and generalization to multiple objects.

Food picking is trivial for humans but not for robots, as foods are fragile. Presetting foods' physical properties does not help robots much due to the objects' inter- and intra-category diversity. A recent study proved that learning-based fracture anticipation with tactile sensors could overcome this problem; however, the method trains the model for each food to deal with intra-category differences, and tuning robots for each food leads to an undesirable amount of food consumption. This study proposes a novel framework for learning food-picking tasks without consuming foods. The key idea is to leverage the object-breaking experiences of several reusable fragile objects instead of consuming real foods while making the picking ability object-invariant with domain generalization (DG). In real-robot experiments, we trained a model with reusable objects (toy blocks, ping-pong balls, and jellies), selected based on the three common fracture types (crack, rupture, and crush). We then tested the model with four real food objects (tofu, bananas, potato chips, and tomatoes). The results showed that the proposed combination of reusable objects' breaking experiences and DG is effective for the food-picking task.

A factored Nonlinear Program (Factored-NLP) explicitly models the dependencies between a set of continuous variables and nonlinear constraints, providing an expressive formulation for relevant robotics problems such as manipulation planning or simultaneous localization and mapping. When the problem is over-constrained or infeasible, a fundamental issue is to detect a minimal subset of variables and constraints that are infeasible. Previous approaches require solving several nonlinear programs, incrementally adding and removing constraints, and are thus computationally expensive. In this paper, we propose a graph neural architecture that predicts which variables and constraints are jointly infeasible. The model is trained with a dataset of labeled subgraphs of Factored-NLPs, and importantly, can make useful predictions on larger factored nonlinear programs than the ones seen during training. We evaluate our approach in robotic manipulation planning, where our model is able to generalize to longer manipulation sequences involving more objects and robots, and different geometric environments. The experiments show that the learned model accelerates general algorithms for conflict extraction (by a factor of 50) and heuristic algorithms that exploit expert knowledge (by a factor of 4).

Tasks involving locally unstable or discontinuous dynamics (such as bifurcations and collisions) remain challenging in robotics, because small variations in the environment can have a significant impact on task outcomes. For such tasks, learning a robust deterministic policy is difficult. We focus on structuring exploration with multiple stochastic policies based on a mixture of experts (MoE) policy representation that can be efficiently adapted. The MoE policy is composed of stochastic sub-policies that allow exploration of multiple distinct regions of the action space (or strategies) and a high-level selection policy to guide exploration towards the most promising regions. We develop a robot system to evaluate our approach in a real-world physical problem solving domain. After training the MoE policy in simulation, online learning in the real world demonstrates efficient adaptation within just a few dozen attempts, with a minimal sim2real gap. Our results confirm that representing multiple strategies promotes efficient adaptation in new environments and strategies learned under different dynamics can still provide useful information about where to look for good strategies.

Learning diverse dexterous manipulation behaviors with assorted objects remains an open grand challenge. While policy learning methods offer a powerful avenue to attack this problem, these approaches require extensive per-task engineering and algorithmic tuning. This paper seeks to escape these constraints, by developing a Pre-Grasp informed Dexterous Manipulation (PGDM) framework that generates diverse dexter-ous manipulation behaviors, without any task-specific reasoning or hyper-parameter tuning. At the core of PGDM is a well known robotics construct, pre-grasps (i.e. the hand-pose preparing for object interaction). This simple primitive is enough to induce efficient exploration strategies for acquiring complex dexterous manipulation behaviors. To exhaustively verify these claims, we introduce TCDM, a benchmark of 50 diverse manipulation tasks defined over multiple objects and dexterous manipulators. Tasks for TCDM are defined automatically using exemplar object trajectories from diverse sources (animators, human behaviors, etc.), without any per-task engineering and/or supervision. Our experiments validate that PGDM's exploration strategy, induced by a surprisingly simple ingredient (single pre-grasp pose), matches the performance of prior methods, which require expensive per-task feature/reward engineering, expert supervision, and hyper-parameter tuning. For animated visualizations, trained policies, and project code, please refer to https://pregrasps.github.io/.

We propose a technique for depth completion of transparent objects using augmented data captured directly from real environments with complicated geometry. Using cyclic adversarial learning we train translators to convert between painted versions of the objects and their real transparent counterpart. The translators are trained on unpaired data, hence datasets can be created rapidly and without any manual labeling. Our technique does not make any assumptions about the geometry of the environment, unlike SOTA systems that assume easily observable occlusion and contact edges, such as ClearGrasp. We show how our technique outperforms ClearGrasp in a dishwasher environment, in which occlusion and contact edges are difficult to observe. We also show how the technique can be used to create an object manipulation application with a humanoid robot. Supplementary URI: https://ftorise.github.io/faking_depth_web/.

3D printing based on continuous deposition of materials, such as filament-based 3D printing, has seen widespread adoption thanks to its versatility in working with a wide range of materials. An important shortcoming of this type of technology is its limited multi-material capabilities. While there are simple hardware designs that enable multi-material printing in principle, the required software is heavily underdeveloped. A typical hardware design fuses together individual materials fed into a single chamber from multiple inlets before they are deposited. This design, however, introduces a time delay between the intended material mixture and its actual deposition. In this work, inspired by diverse path planning research in robotics, we show that this mechanical challenge can be addressed via improved printer control. We propose to formulate the search for optimal multi-material printing policies in a reinforcement learning setup. We put forward a simple numerical deposition model that takes into account the non-linear material mixing and delayed material deposition. To validate our system we focus on color fabrication, a problem known for its strict requirements for varying material mixtures at a high spatial frequency. We demonstrate that our learned control policy outperforms state-of-the-art hand-crafted algorithms.