Humanoid robots are increasingly being deployed in industrial scenarios. However, creating controllers for their locomotion remains challenging, especially in unstructured environments. In this work, we leverage the morphological symmetry about the sagittal plane in high-dimensional humanoid agents to enhance locomotion learning. We incorporate group-equivariant neural networks in reinforcement learning (RL) through Proximal Policy Optimization (PPO). Using Equivariant Multi-Layer Perceptrons (EMLP), we aim to improve sample efficiency and RL training stability. Our experiments with the 21-DoF Unitree H1 humanoid suggest that while EMLP combined with PPO improves sample efficiency, vanilla PPO achieves marginally higher performance in terms of gait quality and biomechanical realism. For assessment, we propose a suite of human-inspired biomechanical metrics, such as joint trajectory deviation, gait symmetry, phase consistency, energy efficiency, and motion smoothness—comparing learned policies against human motion capture (MoCap) data. These metrics are used for quantitatively evaluating the similarity between human gait and reinforcement learning-based humanoid locomotion. Surprisingly, despite the theoretical benefits of equivariance, our findings suggest that excessive symmetry constraints may limit expressivity and impede the emergence of human-like locomotion in complex agents. This study provides key insights into the trade-offs of incorporating symmetry priors in deep reinforcement learning for humanoid control.

Grasping and manipulation remain fundamental challenges in the effective deployment of robotic systems in real-world applications. In retail and supermarket scenarios, soft robotic grippers enable safe and efficient material handling. However, existing grasp planners are designed for rigid or suction-based grippers. Soft grasping is more challenging in terms of planning, estimation and sensing due to deflections in the gripper material on contact with the target. We present a system for soft robotic grasping using a custom gripper with a reconfigurable wrist that leverages reinforcement learning to augment existing vision-based techniques to adapt to the target object’s geometry. This system includes a hidden superquadrics module to guide the adaptation of the gripper’s palm configuration. We evaluate this system in a PyBullet simulation environment and compare it with a baseline synergy-based grasp strategy. Ongoing and future work involves transferring this planner to our physical robotic platform and evaluation in retail stores.

One core component of Active Debris Removal (ADR) and On-Orbit Servicing (OOS) missions in space is to estimate and track the relative pose of a non-cooperative satellite in close proximity. Conventionally, Image Processing methods have been popular in pose estimation by employing manual feature extraction techniques. But the performance of such methods plateaus in the challenging illumination conditions and sensor capability constraints in space, because of which Deep Learning (DL)-based approaches have gained traction. This paper aims to provide an improvement over the existing state-of-the-art direct pose estimation methods from a monocular camera, without relying on any 3D model of the target satellite. The main contribution of this work is to develop a general purpose satellite-invariant pose estimation architecture with improved accuracy and implement an adaptive navigation filter over it to track the pose continuously over a stream of images. The pose estimation module includes a modified DenseNet architecture. In order to test the generalization capability, the proposed pose estimation module is tested on the SPEED, SPEED+, SHIRT and URSO datasets and compared with other existing methods. The advantage of the proposed method is that the same model architecture is able to give accurate pose estimation results for different satellite datasets. To perform continuous tracking of the relative pose, an adaptive EKF (Extended Kalman Filter) is implemented on the initial pose estimates. For performance evaluation of the navigation filter, the accuracy goals required for the relative navigation of Hubble Space Telescope SM4 mission are considered while testing on the SHIRT dataset.

Managing shelves in retail stores includes re-stocking, rearrangement and replenishment of products. As these are some of the most labor-intensive activities, there has been widespread demand from retailers for automation in this domain. However, major challenges still remain in perception, navigation and manipulation while implementing an autonomous robotic system for this purpose. We present a system aimed at addressing some of these challenges through novel approaches. In terms of perception, we have developed a transformer-based local anomaly detection algorithm that can identify misplaced items without the need for a central database. Navigation of the omnidirectional mobile base is performed through stereo vision and LiDAR sensors. Finally, identifying grasping and manipulation as one of the key shortcomings of present robotic systems in this domain, we have developed a customized soft robotic gripper targeted at retail objects. It has compliant cable-driven fingers, and a palm configuration that can be adapted in real-time based on the target object's geometry. Coupled with a conventional two-fingered gripper in a dual-arm setup, this system is equipped to handle most objects encountered in a retail setting. We describe the underlying hardware and algorithms for each component of the system, evaluating their individual performance. We then evaluate the whole system in a mock retail setup, demonstrating promising results for autonomous management of shelves.

The swiftly expanding retail sector is increasingly adopting autonomous mobile robots empowered by artificial intelligence and machine learning algorithms to gain an edge in the competitive market. However, these autonomous robots encounter challenges in adapting to the dynamic nature of retail products, often struggling to operate autonomously in novel situations. In this study, we introduce an omni-directional dual-arm mobile robot specifically tailored for use in retail environments. Additionally, we propose a tele-operation method that enables shared control between the robot and a human operator. This approach utilizes a Virtual Reality (VR) motion capture system to capture the operator's commands, which are then transmitted to the robot located remotely in a retail setting. Furthermore, the robot is equipped with heterogeneous grippers on both manipulators, facilitating the handling of a wide range of items. We validate the efficacy of the proposed system through testing in a mockup of retail environment, demonstrating its ability to manipulate various commonly encountered retail items using both single and dual-arm coordinated manipulation techniques.

Teleoperated robots have enabled humans to manipulate objects in remote environments without requiring physical presence. In this paper we focus on teleoperation of a robotic arm with shared control between the robot and the operator. A model-mediated approach is used to compensate for delays in the communication channel. Position information of the operator's arm is captured and processed to compute the states of a motion prediction model before transmission over a network to be used on the robot's side, allowing for compensation of transmission delays. Model Predictive Control (MPC) and a novel goal prediction algorithm is used to follow the operator's intended motion while reducing the cognitive loads arising from collision avoidance and fine manipulation in the remote environment. We evaluate the proposed method against a baseline pure teleoperation condition with an inverse kinematic controller and observe that the proposed approach improves the overall teleoperation performance in terms of task completion time.

This paper presents a model-free learning-based controller for whole-body control of an orbiting space robot during proximity operations. Proximity control methods for space robots (robotic manipulators mounted on a floating satellite base) enable the robotic arm to reach up to the target body in order to perform autonomous tasks like in-orbit servicing, debris capture, etc. However, coupled motion control of such robots is tricky due to the floating nature of the satellite base. Although conventional controllers have been employed for coupled control of such nonlinear systems, their modeling and sophisticated control become all the more difficult with the increasing degree of freedom of the robot. Model-free Deep Reinforcement Learning (RL) has been successful in learning complex policies in the field of robotic manipulation. However, most of the research in this domain has been focused only on the control of the space robotic arm, and not the satellite base, with a majority of them focusing only on the arm position control. A coupled controller which also simultaneously controls the satellite base orientation is essential for the proper functioning of onboard sensors and equipment which have pointing requirements. This paper uses Proximal Policy Optimization (PPO) algorithm to control the position (3 DOF) and orientation (3 DOF) of the end-effector while also controlling the orientation of the base satellite (3 DOF). To the best knowledge of the authors, a model-free Deep RL method has not yet been used for simultaneous 9 DOF control of a floating space robot so far. We also improvise over the standard reward functions that are used in Deep RL algorithms for improved performance of the learning algorithm. The training of the policy is performed using a PyBullet physics simulator and the comparison of the performance of the learning algorithm against the standard reward functions is presented.