Abstract-- Learning-based methods have proven useful at generating complex motions for robots, including humanoids. Reinforcement learning (RL) has been used to learn locomotion policies, some of which leverage a periodic reward formulation. This work extends the periodic reward formulation of locomotion to skateboarding for the REEM-C robot. Brax/MJX is used to implement the RL problem to achieve fast training. Initial results in simulation are presented with hardware experiments in progress.

Terahertz (THz) space communications (Tera-SpaceCom) is envisioned as a promising technology to enable various space science and communication applications. Mainly, the realm of Tera-SpaceCom consists of THz sensing for space exploration, data centers in space providing cloud services for space exploration tasks, and a low earth orbit (LEO) mega-constellation relaying these tasks to ground stations (GSs) or data centers via THz links. Moreover, to reduce the computational burden on data centers as well as resource consumption and latency in the relaying process, the LEO mega-constellation provides satellite edge computing (SEC) services to directly compute space exploration tasks without relaying these tasks to data centers. The LEO satellites that receive space exploration tasks offload (i.e., distribute) partial tasks to their neighboring LEO satellites, to further reduce their computational burden. However, efficient joint communication resource allocation and computing task offloading for the Tera-SpaceCom SEC network is an NP-hard mixed-integer nonlinear programming problem (MINLP), due to the discrete nature of space exploration tasks and sub-arrays as well as the continuous nature of transmit power. To tackle this challenge, a graph neural network (GNN)-deep reinforcement learning (DRL)-based joint resource allocation and task offloading (GRANT) algorithm is proposed with the target of long-term resource efficiency (RE). Particularly, GNNs learn relationships among different satellites from their connectivity information. Furthermore, multi-agent and multi-task mechanisms cooperatively train task offloading and resource allocation. Compared with benchmark solutions, GRANT not only achieves the highest RE with relatively low latency, but realizes the fewest trainable parameters and the shortest running time.

In this study, we conduct a comprehensive benchmark of the Safe Reinforcement Learning (Safe RL) algorithms for the task of vision-driven river following of Unmanned Aerial Vehicle (UAV) in a Unity-based photo-realistic simulation environment. We empirically validate the effectiveness of semantic-augmented image encoding method, assessing its superiority based on Relative Entropy and the quality of water pixel reconstruction. The determination of the encoding dimension, guided by reconstruction loss, contributes to a more compact state representation, facilitating the training of Safe RL policies. Across all benchmarked Safe RL algorithms, we find that First Order Constrained Optimization in Policy Space achieves the optimal balance between reward acquisition and safety compliance. Notably, our results reveal that on-policy algorithms consistently outperform both off-policy and model-based counterparts in both training and testing environments. Importantly, the benchmarking outcomes and the vision encoding methodology extend beyond UAVs, and are applicable to Autonomous Surface Vehicles (ASVs) engaged in autonomous navigation in confined waters.

Recent studies have shown that cooperative multi-agent deep reinforcement learning (c-MADRL) is under the threat of backdoor attacks. Once a backdoor trigger is observed, it will perform abnormal actions leading to failures or malicious goals. However, existing proposed backdoors suffer from several issues, e.g., fixed visual trigger patterns lack stealthiness, the backdoor is trained or activated by an additional network, or all agents are backdoored. To this end, in this paper, we propose a novel backdoor attack against c-MADRL, which attacks the entire multi-agent team by embedding the backdoor only in a single agent. Firstly, we introduce adversary spatiotemporal behavior patterns as the backdoor trigger rather than manual-injected fixed visual patterns or instant status and control the attack duration. This method can guarantee the stealthiness and practicality of injected backdoors. Secondly, we hack the original reward function of the backdoored agent via reward reverse and unilateral guidance during training to ensure its adverse influence on the entire team. We evaluate our backdoor attacks on two classic c-MADRL algorithms VDN and QMIX, in a popular c-MADRL environment SMAC. The experimental results demonstrate that our backdoor attacks are able to reach a high attack success rate (91.6\%) while maintaining a low clean performance variance rate (3.7\%).

Policy design of non-stationary Markov Decision Processes (MDPs) has always been challenging due to the time-varying system dynamics and rewards, so the learner usually suffers from uncertainties of future rewards and transitions. Fortunately, exogenous predictions are available in many applications. For example, in energy systems, look-ahead information is available in the form of renewable generation forecasts and demand forecasts Amin et al. [2019]. It is intuitive to design an algorithm that controls the energy system by utilizing that information to concentrate energy usage in the time frame with the lowest energy price and lower the overall energy cost. To give another example, smart servers can make predictions of future internet traffic from historical data Katris and Daskalaki [2015]. Given that the server tries to minimize the average waiting time of all tasks, if there is only light traffic, the average waiting time will be most reduced by only using the fastest server. However, if the smart server forecasts that there will be heavy traffic in the future, all servers should work to reduce the length of the queue. However, although policy adaptation in a time-varying environment has been extensively studied [Auer et al., 2008; Richards et al., 2021; Zhang et al., 2024; Gajane et al., 2018], they do not typically take advantage of exogenous predictions.

As a data-driven paradigm, offline reinforcement learning (Offline RL) has been formulated as sequence modeling, where the Decision Transformer (DT) has demonstrated exceptional capabilities. Unlike previous reinforcement learning methods that fit value functions or compute policy gradients, DT adjusts the autoregressive model based on the expected returns, past states, and actions, using a causally masked Transformer to output the optimal action. However, due to the inconsistency between the sampled returns within a single trajectory and the optimal returns across multiple trajectories, it is challenging to set an expected return to output the optimal action and stitch together suboptimal trajectories. Decision ConvFormer (DC) is easier to understand in the context of modeling RL trajectories within a Markov Decision Process compared to DT. We propose the Q-value Regularized Decision ConvFormer (QDC), which combines the understanding of RL trajectories by DC and incorporates a term that maximizes action values using dynamic programming methods during training. This ensures that the expected returns of the sampled actions are consistent with the optimal returns. QDC achieves excellent performance on the D4RL benchmark, outperforming or approaching the optimal level in all tested environments. It particularly demonstrates outstanding competitiveness in trajectory stitching capability.

In particular, modelling equations for different musical structural parameters were defined, namely: rhythmic structure, sound level/perceptual loudness, and tempo, modality, pitch register and pitch contour. Through the creation of two agents, these parameters are incorporated, preserving the original equations in this new mode of generation. As the integration of affective models forms the basis of this exploration, how these human affective states are modelled follows the valence-arousal model introduced by Russell (1980). Formerly, music psychology literature labelled affective states using a categorical model, suggesting these stem from a finite number of monopolar universal basic affects. However, currently various two or three-dimensional models have been more universally adopted, with Russell's circumplex model of affect being commonly used, due to its ability to represent the complexities of affect. This approach employs valence (pleasure vs. displeasure) and arousal (high vs. low energy) as its dimensions, and is used in the research. The reinforcement learning problem in the context of generating musical code based on valence-arousal coordinates involves training an agent to select sequences of code that correspond to desired affective qualities.

Formal verification is a promising method for producing reliable software, but the difficulty of manually writing verification proofs severely limits its utility in practice. Recent methods have automated some proof synthesis by guiding a search through the proof space using a theorem prover. Unfortunately, the theorem prover provides only the crudest estimate of progress, resulting in effectively undirected search. To address this problem, we create QEDCartographer, an automated proof-synthesis tool that combines supervised and reinforcement learning to more effectively explore the proof space. QEDCartographer incorporates the proofs' branching structure, enabling reward-free search and overcoming the sparse reward problem inherent to formal verification. We evaluate QEDCartographer using the CoqGym benchmark of 68.5K theorems from 124 open-source Coq projects. QEDCartographer fully automatically proves 21.4% of the test-set theorems. Previous search-based proof-synthesis tools Tok, Tac, ASTactic, Passport, and Proverbot9001, which rely only on supervised learning, prove 9.6%, 9.8%, 10.9%, 12.5%, and 19.8%, respectively. Diva, which combines 62 tools, proves 19.2%. Comparing to the most effective prior tool, Proverbot9001, QEDCartographer produces 34% shorter proofs 29% faster, on average over the theorems both tools prove. Together, QEDCartographer and non-learning-based CoqHammer prove 30.3% of the theorems, while CoqHammer alone proves 26.6%. Our work demonstrates that reinforcement learning is a fruitful research direction for improving proof-synthesis tools' search mechanisms.

Graph path search is a classic computer science problem that has been recently approached with Reinforcement Learning (RL) due to its potential to outperform prior methods. Existing RL techniques typically assume a global view of the network, which is not suitable for large-scale, dynamic, and privacy-sensitive settings. An area of particular interest is search in social networks due to its numerous applications. Inspired by seminal work in experimental sociology, which showed that decentralized yet efficient search is possible in social networks, we frame the problem as a collaborative task between multiple agents equipped with a limited local view of the network. We propose a multi-agent approach for graph path search that successfully leverages both homophily and structural heterogeneity. Our experiments, carried out over synthetic and real-world social networks, demonstrate that our model significantly outperforms learned and heuristic baselines. Furthermore, our results show that meaningful embeddings for graph navigation can be constructed using reward-driven learning.

We present DemoStart, a novel auto-curriculum reinforcement learning method capable of learning complex manipulation behaviors on an arm equipped with a three-fingered robotic hand, from only a sparse reward and a handful of demonstrations in simulation. Learning from simulation drastically reduces the development cycle of behavior generation, and domain randomization techniques are leveraged to achieve successful zero-shot sim-to-real transfer. Transferred policies are learned directly from raw pixels from multiple cameras and robot proprioception. Our approach outperforms policies learned from demonstrations on the real robot and requires 100 times fewer demonstrations, collected in simulation. More details and videos in https://sites.google.com/view/demostart.