Meta-Reinforcement Learning (Meta-RL) enables fast adaptation to new testing tasks. Despite recent advancements, it is still challenging to learn performant policies across multiple complex and high-dimensional tasks. To address this, we propose a novel architecture with three hierarchical levels for 1) learning task representations, 2) discovering task-agnostic macro-actions in an automated manner, and 3) learning primitive actions. The macro-action can guide the low-level primitive policy learning to more efficiently transition to goal states. This can address the issue that the policy may forget previously learned behavior while learning new, conflicting tasks. Moreover, the task-agnostic nature of the macro-actions is enabled by removing task-specific components from the state space. Hence, this makes them amenable to re-composition across different tasks and leads to promising fast adaptation to new tasks. Also, the prospective instability from the tri-level hierarchies is effectively mitigated by our innovative, independently tailored training schemes. Experiments in the MetaWorld framework demonstrate the improved sample efficiency and success rate of our approach compared to previous state-of-the-art methods.

Neural scaling laws are observed in a range of domains, to date with no clear understanding of why they occur. Recent theories suggest that loss power laws arise from Zipf's law, a power law observed in domains like natural language. One theory suggests that language scaling laws emerge when Zipf-distributed task quanta are learned in descending order of frequency. In this paper we examine power-law scaling in AlphaZero, a reinforcement learning algorithm, using a theory of language-model scaling. We find that game states in training and inference data scale with Zipf's law, which is known to arise from the tree structure of the environment, and examine the correlation between scaling-law and Zipf's-law exponents. In agreement with quanta scaling theory, we find that agents optimize state loss in descending order of frequency, even though this order scales inversely with modelling complexity. We also find that inverse scaling, the failure of models to improve with size, is correlated with unusual Zipf curves where end-game states are among the most frequent states. We show evidence that larger models shift their focus to these less-important states, sacrificing their understanding of important early-game states.

Reinforcement learning (RL) algorithms aim to balance exploiting the current best strategy with exploring new options that could lead to higher rewards. Most common RL algorithms use undirected exploration, i.e., select random sequences of actions. Exploration can also be directed using intrinsic rewards, such as curiosity or model epistemic uncertainty. However, effectively balancing task and intrinsic rewards is challenging and often task-dependent. In this work, we introduce a framework, MaxInfoRL, for balancing intrinsic and extrinsic exploration. MaxInfoRL steers exploration towards informative transitions, by maximizing intrinsic rewards such as the information gain about the underlying task. When combined with Boltzmann exploration, this approach naturally trades off maximization of the value function with that of the entropy over states, rewards, and actions. We show that our approach achieves sublinear regret in the simplified setting of multi-armed bandits. We then apply this general formulation to a variety of off-policy model-free RL methods for continuous state-action spaces, yielding novel algorithms that achieve superior performance across hard exploration problems and complex scenarios such as visual control tasks.

Real-world accidents in learning-enabled CPS frequently occur in challenging corner cases. During the training of deep reinforcement learning (DRL) policy, the standard setup for training conditions is either fixed at a single initial condition or uniformly sampled from the admissible state space. This setup often overlooks the challenging but safety-critical corner cases. To bridge this gap, this paper proposes a physics-model-guided worst-case sampling strategy for training safe policies that can handle safety-critical cases toward guaranteed safety. Furthermore, we integrate the proposed worst-case sampling strategy into the physics-regulated deep reinforcement learning (Phy-DRL) framework to build a more data-efficient and safe learning algorithm for safety-critical CPS. We validate the proposed training strategy with Phy-DRL through extensive experiments on a simulated cart-pole system, a 2D quadrotor, a simulated and a real quadruped robot, showing remarkably improved sampling efficiency to learn more robust safe policies.

Abstract--Reinforcement learning (RL) has shown great effectiveness in quadrotor control, enabling specialized policies to develop even human-champion-level performance in singletask scenarios. To address this limitation, this paper presents a novel multi-task reinforcement learning (MTRL) framework tailored for quadrotor control, leveraging the shared physical dynamics of the platform to enhance sample efficiency and task performance. By employing a multi-critic architecture and shared task encoders, our framework facilitates knowledge transfer across tasks, enabling a single policy to execute diverse maneuvers, including high-speed stabilization, velocity tracking, and autonomous racing. Our experimental results, validated both in simulation and real-world scenarios, demonstrate that our framework outperforms baseline approaches in terms of sample efficiency and overall task performance. EAL world quadrotor applications typically involve multiple tasks and skills.

UAV's are becoming popular for various object search applications in agriculture, however they usually use time-consuming row-by-row flight paths. This paper presents a deep-reinforcement-learning method for path planning to efficiently localize objects of interest using UAVs with a minimal flight-path length. The method uses some global prior knowledge with uncertain object locations and limited resolution in combination with a local object map created using the output of an object detection network. The search policy could be learned using deep Q-learning. We trained the agent in simulation, allowing thorough evaluation of the object distribution, typical errors in the perception system and prior knowledge, and different stopping criteria. When objects were non-uniformly distributed over the field, the agent found the objects quicker than a row-by-row flight path, showing that it learns to exploit the distribution of objects. Detection errors and quality of prior knowledge had only minor effect on the performance, indicating that the learned search policy was robust to errors in the perception system and did not need detailed prior knowledge. Without prior knowledge, the learned policy was still comparable in performance to a row-by-row flight path. Finally, we demonstrated that it is possible to learn the appropriate moment to end the search task. The applicability of the approach for object search on a real drone was comprehensively discussed and evaluated. Overall, we conclude that the learned search policy increased the efficiency of finding objects using a UAV, and can be applied in real-world conditions when the specified assumptions are met.

This paper reviews the large spectrum of methods for generating robot motion proposed over the 50 years of robotics research culminating in recent developments. It crosses the boundaries of methodologies, typically not surveyed together, from those that operate over explicit models to those that learn implicit ones.

Mixed-Integer Linear Programming (MILP) is an optimization technique widely used in various fields. Existing end-to-end learning methods for MILP generate values for a subset of decision variables and delegate the remaining problem to traditional MILP solvers. However, this approach does not guarantee solution feasibility (i.e., satisfying all constraints) due to inaccurate predictions and primarily focuses on prediction for binary decision variables. When addressing MILP involving non-binary integer variables using machine learning (ML), feasibility issues can become even more pronounced. Since finding an optimal solution requires satisfying all constraints, addressing feasibility is critical. To overcome these limitations, we propose a novel reinforcement learning (RL)-based solver that interacts with MILP to incrementally discover better feasible solutions without relying on traditional solvers. We design reward functions tailored for MILP, which enable the RL agent to learn relationships between decision variables and constraints. Furthermore, we leverage a Transformer encoder-based graph neural network (GNN) to effectively model complex relationships among decision variables. Our experimental results demonstrate that the proposed method can solve MILP problems and find near-optimal solutions without delegating the remainder to traditional solvers. The proposed method provides a meaningful step forward as an initial study in solving MILP problems entirely with ML in an end-to-end manner.

Modern society heavily relies on robotic systems, their use affects the aerospace, automotive, energy, disaster response, health care, manufacturing, and traffic management industries among countless others. From making robots walk Westervelt et al. [2007] to getting molecular swarms to kill cancer cells Wijewardhane et al. [2022], whole fields of research dedicate themselves to the problem of control. Intelligently selecting control strategies so that we can manage, direct, or command the trajectories a system can take distills the essence of problems faced in control. When a system can be controlled in the aforementioned manner using control loops, the system in question is termed a control system. Tackling the problem of control, the research community has produced many alternative solutions with varying trade-offs concerning what is achievable and how much we can represent these systems and our goals.

This paper presents CLIP-RLDrive, a new reinforcement learning (RL)-based framework for improving the decision-making of autonomous vehicles (AVs) in complex urban driving scenarios, particularly in unsignalized intersections. To achieve this goal, the decisions for AVs are aligned with human-like preferences through Contrastive Language-Image Pretraining (CLIP)-based reward shaping. One of the primary difficulties in RL scheme is designing a suitable reward model, which can often be challenging to achieve manually due to the complexity of the interactions and the driving scenarios. To deal with this issue, this paper leverages Vision-Language Models (VLMs), particularly CLIP, to build an additional reward model based on visual and textual cues.