Deep reinforcement learning (DRL) has had success in virtual and simulated domains, but due to key differences between simulated and real-world environments, DRL-trained policies have had limited success in real-world applications. To assist researchers to bridge the \textit{sim-to-real gap}, in this paper, we describe a low-cost physical inverted pendulum apparatus and software environment for exploring sim-to-real DRL methods. In particular, the design of our apparatus enables detailed examination of the delays that arise in physical systems when sensing, communicating, learning, inferring and actuating. Moreover, we wish to improve access to educational systems, so our apparatus uses readily available materials and parts to reduce cost and logistical barriers. Our design shows how commercial, off-the-shelf electronics and electromechanical and sensor systems, combined with common metal extrusions, dowel and 3D printed couplings provide a pathway for affordable physical DRL apparatus. The physical apparatus is complemented with a simulated environment implemented using a high-fidelity physics engine and OpenAI Gym interface.

Training robotic manipulation policies traditionally requires numerous demonstrations and/or environmental rollouts. While recent Imitation Learning (IL) and Reinforcement Learning (RL) methods have reduced the number of required demonstrations, they still rely on expert knowledge to collect high-quality data, limiting scalability and accessibility. We propose Sketch-to-Skill, a novel framework that leverages human-drawn 2D sketch trajectories to bootstrap and guide RL for robotic manipulation. Our approach extends beyond previous sketch-based methods, which were primarily focused on imitation learning or policy conditioning, limited to specific trained tasks. Sketch-to-Skill employs a Sketch-to-3D Trajectory Generator that translates 2D sketches into 3D trajectories, which are then used to autonomously collect initial demonstrations. We utilize these sketch-generated demonstrations in two ways: to pre-train an initial policy through behavior cloning and to refine this policy through RL with guided exploration. Experimental results demonstrate that Sketch-to-Skill achieves ~96% of the performance of the baseline model that leverages teleoperated demonstration data, while exceeding the performance of a pure reinforcement learning policy by ~170%, only from sketch inputs. This makes robotic manipulation learning more accessible and potentially broadens its applications across various domains.

In cooperative multi-agent reinforcement learning (MARL), the permutation problem where the state space grows exponentially with the number of agents reduces sample efficiency. Additionally, many existing architectures struggle with scalability, relying on a fixed structure tied to a specific number of agents, limiting their applicability to environments with a variable number of entities. While approaches such as graph neural networks (GNNs) and self-attention mechanisms have progressed in addressing these challenges, they have significant limitations as dense GNNs and self-attention mechanisms incur high computational costs. To overcome these limitations, we propose a novel agent network and a non-linear mixing network that ensure permutation-equivariance and scalability, allowing them to generalize to environments with various numbers of agents. Our agent network significantly reduces computational complexity, and our scalable hypernetwork enables efficient weight generation for non-linear mixing. Additionally, we introduce curriculum learning to improve training efficiency. Experiments on SMACv2 and Google Research Football (GRF) demonstrate that our approach achieves superior learning performance compared to existing methods. By addressing both permutation-invariance and scalability in MARL, our work provides a more efficient and adaptable framework for cooperative MARL. Our code is available at https://github.com/funny-rl/SPECTra.

Large language models memorize sensitive data from their pretraining corpora Jang et al. (2023). In this work, we propose CE-U (Cross Entropy Unlearning), a loss function for unlearning. CE-U addresses fundamental limitations of gradient ascent approaches that suffer from vanishing gradients when model confidence is high and exploding gradients when confidence is low. We also unify standard cross entropy learning and unlearning into a single framework. On the TOFU benchmark for unlearning Maini et al. (2024), CE-U achieves state-of-the-art results on LLaMA2-7B models without using an extra oracle model or additional positive samples. Our analysis reveals that the problematic gradient ascent component also exists in reinforcement learning algorithms like DPO Rafailov et al. (2023) and GRPO Shao et al. (2024). This suggests that applying CE-U approach to reinforcement learning could be promising to improve stability and convergence.

We present heuristically optimal strategies expressed by deep learning agents playing a simple avoidance game. We analyse the learning and behaviour of two agents within a symmetrical grid world that must cross paths to reach a target destination without crashing into each other or straying off of the grid world in the wrong direction. The agent policy is determined by one neural network that is employed in both agents. Our findings indicate that the fully trained network exhibits behaviour similar to that of the game Hawks and Doves, in that one agent employs an aggressive strategy to reach the target while the other learns how to avoid the aggressive agent.

Target search problems are central to a wide range of fields, from biological foraging to the optimization algorithms. Recently, the ability to reset the search has been shown to significantly improve the searcher's efficiency. However, the optimal resetting strategy depends on the specific properties of the search problem and can often be challenging to determine. In this work, we propose a reinforcement learning (RL)-based framework to train agents capable of optimizing their search efficiency in environments by learning how to reset. First, we validate the approach in a well-established benchmark: the Brownian search with resetting. There, RL agents consistently recover strategies closely resembling the sharp resetting distribution, known to be optimal in this scenario. We then extend the framework by allowing agents to control not only when to reset, but also their spatial dynamics through turning actions. In this more complex setting, the agents discover strategies that adapt both resetting and turning to the properties of the environment, outperforming the proposed benchmarks. These results demonstrate how reinforcement learning can serve both as an optimization tool and a mechanism for uncovering new, interpretable strategies in stochastic search processes with resetting.

Q-learning is a widely used reinforcement learning (RL) algorithm for optimizing wireless networks, but faces challenges with large state-spaces. Recently proposed multi-environment mixed Q-learning (MEMQ) algorithm addresses these challenges by employing multiple Q-learning algorithms across multiple synthetically generated, distinct but structurally related environments, so-called digital cousins. In this paper, we propose a novel multi-agent MEMQ (M-MEMQ) for cooperative decentralized wireless networks with multiple networked transmitters (TXs) and base stations (BSs). TXs do not have access to global information (joint state and actions). The new concept of coordinated and uncoordinated states is introduced. In uncoordinated states, TXs act independently to minimize their individual costs and update local Q-functions. In coordinated states, TXs use a Bayesian approach to estimate the joint state and update the joint Q-functions. The cost of information-sharing scales linearly with the number of TXs and is independent of the joint state-action space size. Several theoretical guarantees, including deterministic and probabilistic convergence, bounds on estimation error variance, and the probability of misdetecting the joint states, are given. Numerical simulations show that M-MEMQ outperforms several decentralized and centralized training with decentralized execution (CTDE) multi-agent RL algorithms by achieving 55% lower average policy error (APE), 35% faster convergence, 50% reduced runtime complexity, and 45% less sample complexity. Furthermore, M-MEMQ achieves comparable APE with significantly lower complexity than centralized methods. Simulations validate the theoretical analyses.

Uncertainty quantification is a critical aspect of reinforcement learning and deep learning, with numerous applications ranging from efficient exploration and stable offline reinforcement learning to outlier detection in medical diagnostics. The scale of modern neural networks, however, complicates the use of many theoretically well-motivated approaches such as full Bayesian inference. Approximate methods like deep ensembles can provide reliable uncertainty estimates but still remain computationally expensive. In this work, we propose contextual similarity distillation, a novel approach that explicitly estimates the variance of an ensemble of deep neural networks with a single model, without ever learning or evaluating such an ensemble in the first place. Our method builds on the predictable learning dynamics of wide neural networks, governed by the neural tangent kernel, to derive an efficient approximation of the predictive variance of an infinite ensemble. Specifically, we reinterpret the computation of ensemble variance as a supervised regression problem with kernel similarities as regression targets. The resulting model can estimate predictive variance at inference time with a single forward pass, and can make use of unlabeled target-domain data or data augmentations to refine its uncertainty estimates. We empirically validate our method across a variety of out-of-distribution detection benchmarks and sparse-reward reinforcement learning environments. We find that our single-model method performs competitively and sometimes superior to ensemble-based baselines and serves as a reliable signal for efficient exploration. These results, we believe, position contextual similarity distillation as a principled and scalable alternative for uncertainty quantification in reinforcement learning and general deep learning.

Group Relative Policy Optimization (GRPO) was introduced and used successfully to train DeepSeek R1 models for promoting reasoning capabilities of LLMs using verifiable or binary rewards. We show in this paper that GRPO with verifiable rewards can be written as a Kullback Leibler ($\mathsf{KL}$) regularized contrastive loss, where the contrastive samples are synthetic data sampled from the old policy. The optimal GRPO policy $\pi_{n}$ can be expressed explicitly in terms of the binary reward, as well as the first and second order statistics of the old policy ($\pi_{n-1}$) and the reference policy $\pi_0$. Iterating this scheme, we obtain a sequence of policies $\pi_{n}$ for which we can quantify the probability of success $p_n$. We show that the probability of success of the policy satisfies a recurrence that converges to a fixed point of a function that depends on the initial probability of success $p_0$ and the regularization parameter $\beta$ of the $\mathsf{KL}$ regularizer. We show that the fixed point $p^*$ is guaranteed to be larger than $p_0$, thereby demonstrating that GRPO effectively amplifies the probability of success of the policy.

This paper presents a fully autonomous robotic system that performs sim-to-real transfer in complex long-horizon tasks involving navigation, recognition, grasping, and stacking in an environment with multiple obstacles. The key feature of the system is the ability to overcome typical sensing and actuation discrepancies during sim-to-real transfer and to achieve consistent performance without any algorithmic modifications. To accomplish this, a lightweight noise-resistant visual perception system and a nonlinearity-robust servo system are adopted. We conduct a series of tests in both simulated and real-world environments. The visual perception system achieves the speed of 11 ms per frame due to its lightweight nature, and the servo system achieves sub-centimeter accuracy with the proposed controller. Both exhibit high consistency during sim-to-real transfer. Benefiting from these, our robotic system took first place in the mineral searching task of the Robotic Sim2Real Challenge hosted at ICRA 2024.