IROS 2022 Papers

Cloud robotics, i.e., controlling robots from the cloud, make it possible to perform more complex processes, make robots smaller, and coordinate multi-robots by sharing information between robots and utilizing abundant computing resources. In cloud robotics, robots need to transmit videos to the cloud in real time to recognize their surroundings. Lowering the video quality reduces the bitrate in low bandwidth environments; however, this may lead to control errors and misrecognition due to lack of detailed image features. Even with 5G, bandwidth fluctuates widely, especially in moving robots, making it difficult to upload high quality video consistently. To reduce bitrate while preserving Quality of Control (QoC), we propose a method of learning the important regions for a pretrained autonomous agent using self-attention, and transmitting the video to the agent by controlling the image quality of each region on the basis of the estimated importance. The evaluation results demonstrate that our approach can maintain QoC while reducing the bitrate to 26% by setting important regions to high quality and the rest to low quality.

Motion optimization is an effective framework for generating smooth and safe trajectories for robotic manipulation tasks. However, it suffers from local optima that hinder its applicability, especially for multi-objective tasks. In this paper, we study this problem in light of the integration of Energy-Based Models (EBM) as guiding priors in motion optimization. EBMs are probabilistic models with unnormalized energy functions that represent expressive multimodal distributions. Due to their implicit nature, EBMs can easily be integrated as data-driven factors or initial sampling distributions in the motion optimization problem. This work presents a set of necessary modeling and algorithmic choices to effectively learn and integrate EBMs into motion optimization. We present a set of EBM architectures for learning generalizable distributions over trajectories that are important for the subsequent deployment of EBMs. Moreover, we investigate the benefit of including smoothness regularization in the learning process to improve motion optimization. In addition to gradient-based solvers, we also propose a stochastic method for trajectory optimization with learned EBMs. We provide extensive empirical results in a set of representative tasks against competitive baselines that demonstrate the superiority of EBMs as priors in motion optimization scaling up to 7 -dof robot pouring that can be easily transferred to the real robotic system. Videos and additional details are available at https://sites.google.com/view/implicit-priors.

Learning goal conditioned control in the real world is a challenging open problem in robotics. Reinforcement learning systems have the potential to learn autonomously via trial-and-error, but in practice the costs of manual reward design, ensuring safe exploration, and hyperparameter tuning are often enough to preclude real world deployment. Imitation learning approaches, on the other hand, offer a simple way to learn control in the real world, but typically require costly cu-rated demonstration data and lack a mechanism for continuous improvement. Recently, iterative imitation methods have been shown to be effective at relaxing both these constraints, learning goal directed control from undirected demonstration data, and improving continuously via self-supervised goal reaching. These approaches, however, have not yet been shown to scale beyond simple simulated environments. In this work, we present the first evidence that simple iterative imitation learning can scale to goal-directed behavior on a real robot in a dynamic setting: high speed, precision table tennis (e.g. “land the ball on this particular target”). We find that this approach offers a straightforward way to do continuous on-robot learning, without complexities such as reward design, value function learning, or sim-to-real transfer. We also find that this approach is scalable-sample efficient enough to train on a physical robot in just a few hours. In real world evaluations, we find that that the resulting policy can perform on par or better than amateur humans (with players sampled randomly from a robotics lab) at the task of returning the ball to specific targets on the table. Finally, we analyze the effect of an initial undirected bootstrap dataset size on performance, finding that a modest amount of unstructured demonstration data provided up-front drastically speeds up the convergence of a general purpose goal-reaching policy. See supplementary video for examples of the policy on a physical robot. Show More

Pushing objects through cluttered scenes is a challenging task, especially when the objects to be pushed have initially unknown dynamics and touching other entities has to be avoided to reduce the risk of damage. In this paper, we approach this problem by applying deep reinforcement learning to generate pushing actions for a robotic manipulator acting on a planar surface where objects have to be pushed to goal locations while avoiding other items in the same workspace. With the latent space learned from a depth image of the scene and other observations of the environment, such as contact information between the end effector and the object as well as distance to the goal, our framework is able to learn contact-rich pushing actions that avoid collisions with other objects. As the experimental results with a six degrees of freedom robotic arm show, our system is able to successfully push objects from start to end positions while avoiding nearby objects. Furthermore, we evaluate our learned policy in comparison to a state-of-the-art pushing controller for mobile robots and show that our agent performs better in terms of success rate, collisions with other objects, and continuous object contact in various scenarios.

Monocular depth estimation (MDE) has attracted intense study due to its low cost and critical functions for robotic tasks such as localization, mapping and obstacle detection. Supervised approaches have led to great success with the advance of deep learning, but they rely on large quantities of ground-truth depth annotations that are expensive to acquire. Unsupervised domain adaptation (UDA) transfers knowledge from labeled source data to unlabeled target data, so as to relax the constraint of supervised learning. However, existing UDA approaches may not completely align the domain gap across different datasets because of the domain shift problem. We believe better domain alignment can be achieved via well-designed feature decomposition. In this paper, we propose a novel UDA method for MDE, referred to as Learning Feature Decomposition for Adaptation (LFDA), which learns to decompose the feature space into content and style components. LFDA only attempts to align the content component since it has a smaller domain gap. Meanwhile, it excludes the style component which is specific to the source domain from training the primary task. Furthermore, LFDA uses separate feature distribution estimations to further bridge the domain gap. Extensive experiments on three domain adaptative MDE scenarios show that the proposed method achieves superior accuracy and lower computational cost compared to the state-of-the-art approaches.

This paper addresses a safe planning and control problem for mobile robots operating in communication- and sensor-limited dynamic environments. In this case the robots cannot sense the objects around them and must instead rely on intermittent, external information about the environment, as e.g., in underwater applications. The challenge in this case is that the robots must plan using only this stale data, while accounting for any noise in the data or uncertainty in the environment. To address this challenge we propose a compositional technique which leverages neural networks to quickly plan and control a robot through crowded and dynamic environments using only intermittent information. Specifically, our tool uses reachability analysis and potential fields to train a neural network that is capable of generating safe control actions. We demonstrate our technique both in simulation with an underwater vehicle crossing a crowded shipping channel and with real experiments with ground vehicles in communication-and sensor-limited environments.

In this work, we propose a learning approach for 3D dynamic bipedal walking when footsteps are constrained to stepping stones. While recent work has shown progress on this problem, real-world demonstrations have been limited to relatively simple open-loop, perception-free scenarios. Our main contribution is a more advanced learning approach that enables real-world demonstrations, using the Cassie robot, of closed-loop dynamic walking over moderately difficult stepping-stone patterns. Our approach first uses reinforcement learning (RL) in simulation to train a controller that maps footstep commands onto joint actions without any reference motion information. We then learn a model of that controller's capabilities, which enables prediction of feasible footsteps given the robot's current dynamic state. The resulting controller and model are then integrated with a real-time overhead camera system for detecting stepping stone locations. For evaluation, we develop a benchmark set of stepping stone patterns, which are used to test performance in both simulation and the real world. Overall, we demonstrate that sim-to-real learning is extremely promising for enabling dynamic locomotion over stepping stones. We also identify challenges remaining that motivate important future research directions.

As human-robot collaboration increases in the workforce, it becomes essential for human-robot teams to coordinate efficiently and intuitively. Traditional approaches for human-robot scheduling either utilize exact methods that are intractable for large-scale problems and struggle to account for stochastic, time varying human task performance, or application-specific heuristics that require expert domain knowledge to develop. We propose a deep learning-based framework, called HybridNet, combining a heterogeneous graph-based encoder with a recurrent schedule propagator for scheduling stochastic human-robot teams under upper- and lower-bound temporal constraints. The HybridNet's encoder leverages Heterogeneous Graph Attention Networks to model the initial environment and team dynamics while accounting for the constraints. By formulating task scheduling as a sequential decision-making process, the HybridNet's recurrent neural schedule propagator leverages Long Short-Term Memory (LSTM) models to propagate forward consequences of actions to carry out fast schedule generation, removing the need to interact with the environment between every taskagent pair selection. The resulting scheduling policy network provides a computationally lightweight yet highly expressive model that is end-to-end trainable via Reinforcement Learning algorithms. We develop a virtual task scheduling environment for mixed human-robot teams in a multi-round setting, capable of modeling the stochastic learning behaviors of human workers. Experimental results showed that HybridNet outperformed other human-robot scheduling solutions across problem sizes for both deterministic and stochastic human performance, with faster runtime compared to pure-GNN-based schedulers.

We propose a simple imitation learning procedure for learning locomotion controllers that can walk over very challenging terrains. We use trajectory optimization (TO) to produce a large dataset of trajectories over procedurally generated terrains and use Reinforcement Learning (RL) to imitate these trajectories. We demonstrate with a realistic model of the ANYmal robot that the learned controllers transfer to unseen terrains and provide an effective initialization for fine-tuning on challenging terrains that require exteroception and precise foot placements. Our setup combines TO and RL in a simple fashion that overcomes the computational limitations and need for a robust tracking controller of the former and the exploration and reward-tuning difficulties of the latter.

Humanoid robots are versatile platforms with the potential for multiple locomotion skills. However, this contact-switched system with only two contact feet is fragile to keep balance in many scenarios. Inspired by birds combining legs and wings, we propose the novel hybrid locomotion behavior for the humanoid robots with the aid of a thruster suit. To fully leverage their agility while guaranteeing efficient computation, we combine the neural controller based on reinforcement learning to handle the complexity of the highly non-linear system and the optimization-based controller to explicitly handle the constraint conditions of the safety-critical thruster module. Our learning framework is demonstrated on several thruster-aided humanoid platforms with hybrid walking and even dynamic locomotion skills. To our best knowledge, it is the first work that, 1. demonstrates agile hybrid whole-body locomotion skills on the thruster-aided humanoid robot; 2. achieves hybrid locomotion under the reinforcement learning settings.

Currently, task-oriented grasp detection approaches are mostly based on pixel-level affordance detection and semantic segmentation. These pixel-level approaches heavily rely on the accuracy of a 2D affordance mask, and the generated grasp candidates are restricted to a small workspace. To mitigate these limitations, we firstly construct a novel affordance-based grasp dataset and propose a 6-DoF task-oriented grasp detection framework, which takes the observed object point cloud as input and predicts diverse 6-DoF grasp poses for different tasks. Specifically, our implicit estimation network and visual affordance network in this framework could directly predict coarse grasp candidates, and corresponding 3D affordance heatmap for each potential task, respectively. Furthermore, the grasping scores from coarse grasps are combined with heatmap values to generate more accurate and finer candidates. Our proposed framework shows significant improvements compared to baselines for existing and novel objects on our simulation dataset. Although our framework is trained based on the simulated objects and environment, the final generated grasp candidates can be accurately and stably executed in the real robot experiments when the object is randomly placed on a support surface.

3D imaging radar offers robust perception capability through visually demanding environments due to the unique penetrative and reflective properties of millimeter waves (mmWave). Current approaches for 3D perception with imaging radar require knowledge of environment geometry, accumulation of data from multiple frames for perception, or access to between-frame motion. Imaging radar presents an additional difficulty due to the complexity of its data representation. To address these issues, and make imaging radar easier to use for downstream robotics tasks, we propose a learning-based method that regresses radar measurements into cylindrical depth maps using LiDAR supervision. Due to the limitation of the regression formulation, directions where the radar beam could not reach will still generate a valid depth. To address this issue, our method additionally learns a 3D filter to remove those pixels. Experiments show that our system generates visually accurate depth estimation. Furthermore, we confirm the overall ability to generalize in the indoor scene using the estimated depth for probabilistic occupancy mapping with ground truth trajectory. The code and model will be released11https://github.com/rpl-cmu/learned-depth-imaging-radar.

Indoor localization systems often fuse inertial odometry with map information via hand-defined methods to reduce odometry drift, but such methods are sensitive to noise and struggle to generalize across odometry sources. To address the robustness problem in map utilization, we propose a data-driven prior on possible user locations in a map by combining learned spatial map embeddings and temporal odometry embeddings. Our prior learns to encode which map regions are feasible locations for a user more accurately than previous hand-defined methods. This prior leads to a 49% improvement in inertial-only localization accuracy when used in a particle filter. This result is significant, as it shows that our relative positioning method can match the performance of absolute positioning using bluetooth beacons. To show the gen-eralizability of our method, we also show similar improvements using wheel encoder odometry. Our code will be made publicly available†1project page: https://rebrand.ly/learned-map-prior.

Bin picking is a challenging problem in robotics due to high dimensional action space, partially visible objects, and contact-rich environments. State-of-the-art methods for bin picking are often simplified as planar manipulation, or learn policy based on human demonstration and motion primitives. The designs have escalated in complexity while still failing to reach the generality and robustness of human picking ability. Here, we present an end-to-end reinforcement learning (RL) framework to produce an adaptable and robust policy for picking objects in diverse real-world environments, including but not limited to tilted bins and corner objects. We present a novel solution to incorporate object interaction in policy learning. The object interaction is represented by the poses of objects. The policy learning is based on two neural networks with asymmetric state inputs. One acts on the object interaction information, while the other acts on the depth observation and proprioceptive signals of robots. The results of experiment shows remarkable zero-shot generalization from simulation to the real world and extensive real-world experiments show the effectiveness of the approach.

We present an incremental search algorithm, called Lifelong-GLS, which combines the vertex efficiency of Lifelong Planning A* (LPA*) and the edge efficiency of Generalized Lazy Search (GLS) for efficient replanning on dynamic graphs where edge evaluation is expensive. We use a lazily evaluated LPA* to repair the cost-to-come inconsistencies of the relevant region of the current search tree based on the previous search results, and then we restrict the expensive edge evaluations only to the current shortest subpath as in the GLS framework. The proposed algorithm is complete and correct in finding the optimal solution in the current graph, if one exists. We also show the efficiency of the proposed algorithm compared to the standard LPA* and the GLS algorithms over consecutive search episodes in a dynamic environment.

There is an increasing demand for robots that can be substituted for humans in various tasks. Mobile robots are being introduced in factories, stores, and public facilities for carrying goods and cleaning. In factories and stores, desks and shelves are arranged such that the work and movement of personnel are reduced. The surrounding furniture is also set to ensure that a single task can be performed in the same place. It is essential to study the intelligence of robots using information from such layouts, wherein human labor and movements are optimized. However, There is no method of map construction or location estimation that uses the characteristics of furniture arrangements that facilitate human work in a work space. Therefore, this study proposes a method for object mapping using layouts in crowded workspaces. Graphically represent the characteristics of furniture placement that make it easy for people to work in a workspace. The links in the graph represent the connections between the objects in the layout property. The nodes are the objects, and the weights of the links represent the strength of the layout properties. This graph is optimized by GraphSLAM to construct a map that considers the arrangement's characteristics. Using the graph structure improves the map's accuracy while allowing for relative changes in placement. The results show a 50.44% improvement in accuracy in a space with 18 desks, followed by two variations of similar desk layouts. The same improvement in accuracy was also observed when the relative positioning of objects changed significantly in each variation, such as a change to the left or right on the same side.

This paper considers the problem of designing legged robots for traversing uneven terrain, wherein terrain characteristics represent uncertainty for the design process. When this process encompasses a wider variety of terrains, the likelihood of the designed robot falling in the real world should decrease. However, computational scalability limits the number of terrains that can be taken into account during design. The proposed framework uses the Alternating Direction Method of Multipliers (ADMM) to solve large-scale concurrent design (co-design) problems. The ADMM coordinates the solution of small-size sub-problems and enforces constraints to reach a consensus on the best design. The framework uses stochastic programming (SP) to account for terrain uncertainty and trajectory optimization (TO) to co-optimize a nominal trajectory alongside hardware parameters and a feedback controller. Case studies demonstrate application for a monopod and a quadruped. For the monopod, ADMM facilitated an increase in the number of terrains considered within co-design by 400% compared to SP alone, which contributed to robustifying the design and decreasing its failure probability to under 1% in an anticipated operating space. A multi-scenario co-design implementation for the quadruped had previously been intractable due to scalability limitations. The ADMM framework, by contrast, shows tractability running with 30 terrain types, opening the horizon for designing more complex systems.

The semantic segmentation of surgical scenes is a prerequisite for task automation in robot assisted interventions. We propose LapSeg3D, a novel DNN-based approach for the voxel-wise annotation of point clouds representing surgical scenes. As the manual annotation of training data is highly time consuming, we introduce a semi-autonomous clustering-based pipeline for the annotation of the gallbladder, which is used to generate segmented labels for the DNN. When evaluated against manually annotated data, LapSeg3D achieves an F1 score of 0.94 for gallbladder segmentation on various datasets of ex-vivo porcine livers. We show LapSeg3D to generalize accurately across different gallbladders and datasets recorded with different RGB-D camera systems.

Autonomous Driving (AD) related features represent important elements for the next generation of mobile robots and autonomous vehicles focused on increasingly intelligent, autonomous, and interconnected systems. The applications involving the use of these features must provide, by definition, real-time decisions, and this property is key to avoid catastrophic accidents. Moreover, all the decision processes must require low power consumption, to increase the lifetime and autonomy of battery-driven systems. These challenges can be addressed through efficient implementations of Spiking Neural Networks (SNNs) on Neuromorphic Chips and the use of event-based cameras instead of traditional frame-based cameras. In this paper, we present a new SNN-based approach, called LaneSNN, for detecting the lanes marked on the streets using the event-based camera input. We develop four novel SNN models characterized by low complexity and fast response, and train them using an offline supervised learning rule. Afterward, we implement and map the learned SNNs models onto the Intel Loihi Neuromorphic Research Chip. For the loss function, we develop a novel method based on the linear composition of Weighted binary Cross Entropy (WCE) and Mean Squared Error (MSE) measures. Our experimental results show a maximum Intersection over Union (IoU) measure of about 0.62 and very low power consumption of about 1 W. The best IoU is achieved with an SNN implementation that occupies only 36 neurocores on the Loihi processor while providing a low latency of less than 8 ms to recognize an image, thereby enabling real-time performance. The IoU measures provided by our networks are comparable with the state-of-the-art, but at a much low power consumption of 1 W.