We introduce Efficient Cross-Episodic Transformers (ECET), a new algorithm for online Meta-Reinforcement Learning that addresses the challenge of enabling reinforcement learning agents to perform effectively in previously unseen tasks. We demonstrate how past episodes serve as a rich source of in-context information, which our model effectively distills and applies to new contexts. Our learned algorithm is capable of outperforming the previous state-of-the-art and provides more efficient meta-training while significantly improving generalization capabilities. Experimental results, obtained across various simulated tasks of the MuJoCo, Meta-World and ManiSkill benchmarks, indicate a significant improvement in learning efficiency and adaptability compared to the state-of-the-art. Our approach enhances the agent's ability to generalize from limited data and paves the way for more robust and versatile AI systems.

In this paper, we propose VLM-Vac, a novel framework designed to enhance the autonomy of smart robot vacuum cleaners. Our approach integrates the zero-shot object detection capabilities of a Vision-Language Model (VLM) with a Knowledge Distillation (KD) strategy. By leveraging the VLM, the robot can categorize objects into actionable classes-either to avoid or to suck-across diverse backgrounds. However, frequently querying the VLM is computationally expensive and impractical for real-world deployment. To address this issue, we implement a KD process that gradually transfers the essential knowledge of the VLM to a smaller, more efficient model. Our real-world experiments demonstrate that this smaller model progressively learns from the VLM and requires significantly fewer queries over time. Additionally, we tackle the challenge of continual learning in dynamic home environments by exploiting a novel experience replay method based on languageguided sampling. Our results show that this approach not only reduces energy consumption by 53 % compared to cumulative learning but also surpasses conventional vision-based clustering methods, particularly in detecting small objects across diverse backgrounds.

Vision-Language-Action Models (VLAs) have demonstrated remarkable generalization capabilities in real-world experiments. However, their success rates are often not on par with expert policies, and they require fine-tuning when the setup changes. In this work, we introduce Refined Policy Distillation (RPD), a novel Reinforcement Learning (RL)-based policy refinement method that bridges this performance gap through a combination of on-policy RL with behavioral cloning. The core idea of RPD is to distill and refine VLAs into compact, high-performing expert policies by guiding the student policy during RL exploration using the actions of a teacher VLA, resulting in increased sample efficiency and faster convergence. We complement our method by fine-tuned versions of Octo and OpenVLA for ManiSkill3 to evaluate RPD in simulation. While this is a key requirement for applying RL, it also yields new insights beyond existing studies on VLA performance in real-world settings. Our experimental results across various manipulation tasks show that RPD enables the RL student to learn expert policies that outperform the VLA teacher in both dense and sparse reward settings, while also achieving faster convergence than the RL baseline. Our approach is even robust to changes in camera perspective and can generalize to task variations that the underlying VLA cannot solve. Our code, dataset, VLA checkpoints, and videos are available at https://refined-policy-distillation.github.io

Reinforcement Learning (RL) has achieved tremendous success in many complex decision-making tasks. However, safety concerns are raised during deploying RL in real-world applications, leading to a growing demand for safe RL algorithms, such as in autonomous driving and robotics scenarios. While safe control has a long history, the study of safe RL algorithms is still in the early stages. To establish a good foundation for future safe RL research, in this paper, we provide a review of safe RL from the perspectives of methods, theories, and applications. First, we review the progress of safe RL from five dimensions and come up with five crucial problems for safe RL being deployed in real-world applications, coined as "2H3W". Second, we analyze the algorithm and theory progress from the perspectives of answering the "2H3W" problems. Particularly, the sample complexity of safe RL algorithms is reviewed and discussed, followed by an introduction to the applications and benchmarks of safe RL algorithms. Finally, we open the discussion of the challenging problems in safe RL, hoping to inspire future research on this thread. To advance the study of safe RL algorithms, we release an open-sourced repository containing major safe RL algorithms at the link.

Object detection is a key technology for intelligent transportation systems (ITSs) to recognize surrounding vehicles. Robust and efficient object detection with roadside sensors could make them more sustainable. This research uses the CARLA simulator to generate synthetic datasets from roadside event-based cameras with multiple weather conditions and evaluates Spiking Neural Networks (SNNs) to improve the sustainability with these datasets. Event-based cameras can detect the change of each pixel intensity asynchronously even under adverse environments such as night. In addition, SNNs have lower energy consumption with neuromorphic hardware than conventional CNNs and can process time-continuous data including event-based data. Evaluations in this research indicate that fine-tuning of YOLOv5 with accumulated event images improves the robustness against adverse weather conditions and SNNs with raw event-based datasets reduce both energy consumption and computational time. Furthermore, the event polarities made object detection more robust against the motion direction of vehicles.

In nature, the movement of quadrupeds is completed under the combined action of the spine and the legs. Inspired by this, this paper explores the effect of a lateral flexing spine on the locomotion of a rat robot. Benefiting from the regular lateral flexion of a soft actuated spine, the rat robot exhibits enhance step length of its hind legs and increased translational velocity by coordinating the opposite movements of the left and right sides. Furthermore, this paper introduces a mathematical model of the effect of the flexible spine on the robot velocity. Finally, extensive experiments are conducted in simulations and on the physical rat robot. Compared with the locomotion without a flexing spine, the simulation results show that the velocity of the robot can be increased up to 218.29%, which is in line with the theoretical results from the proposed mathematical model. Limited by the gap between simulation and the real world, the experiment results of the physical rat robot show a slight performance than the theoretical results. But the physical rat robot can still enhance its translational velocity with the help of a lateral flexing spine.

As a vital cognitive function of animals, the navigation skill is first built on the accurate perception of the directional heading in the environment. Head direction cells (HDCs), found in the limbic system of animals, are proven to play an important role in identifying the directional heading allocentrically in the horizontal plane, independent of the animal's location and the ambient conditions of the environment. However, practical HDC models that can be implemented in robotic applications are rarely investigated, especially those that are biologically plausible and yet applicable to the real world. In this article, we propose a computational HDC network that is consistent with several neurophysiological findings concerning biological HDCs and then implement it in robotic navigation tasks. The HDC network keeps a representation of the directional heading only relying on the angular velocity as an input. We examine the proposed HDC model in extensive simulations and real-world experiments and demonstrate its excellent performance in terms of accuracy and real-time capability.

Abstract Legged locomotion is a challenging task in the field of robotics but a rather simple one in nature. This motivates the use of biological methodologies as solutions to this problem. Central pattern generators are neural networks that are thought to be responsible for locomotion in humans and some animal species. As for robotics, many attempts were made to reproduce such systems and use them for a similar goal. One interesting design model is based on spiking neural networks. This model is the main focus of this work, as its contribution is not limited to engineering but also applicable to neuroscience. This paper introduces a new general framework for building central pattern generators that are task-independent, biologically plausible, and rely on learning methods. The abilities and properties of the presented approach are not only evaluated in simulation but also in a robotic experiment. The results are very promising as the used robot was able to perform stable walking at different speeds and to change speed within the same gait cycle.

Hierarchical Temporal Memory (HTM) is a computational model of the neocortex that is capable of online learning to predict and detect anomalies from continuous data streams. To make HTM also available on power-constrained robot systems, we investigate the feasibility of implementing the model on SpiNNaker, a fully programmable energy-efficient neuromorphic many core system. Our contribution is twofold: First, we propose a mapping of the HTM model components to the SpiNNaker chip architecture. Second, a prototypic implementation of this mapping is successfully evaluated for different sets of model parameters.