Learning global control of underactuated systems with Model-Based Reinforcement Learning Niccol ` o Turcato 1, Marco Cal ` ı 1, Alberto Dalla Libera 1, Giulio Giacomuzzo 1, Ruggero Carli 1 and Diego Romeres 2 Abstract -- This short paper describes our proposed solution for the third edition of the "AI Olympics with RealAIGym" competition, held at ICRA 2025. We employed Monte-Carlo Probabilistic Inference for Learning Control (MC-PILCO), an MBRL algorithm recognized for its exceptional data efficiency across various low-dimensional robotic tasks, including cart-pole, ball & plate, and Furuta pendulum systems. This approach has proven highly effective in physical systems, offering greater data efficiency than Model-Free (MF) alternatives. Notably, MC-PILCO has previously won the first two editions of this competition, demonstrating its robustness in both simulated and real-world environments. Besides briefly reviewing the algorithm, we discuss the most critical aspects of the MC-PILCO implementation in the tasks at hand: learning a global policy for the pendubot and acrobot systems.

Hyperparameter optimisation (HPO) is crucial for achieving strong performance in reinforcement learning (RL), as RL algorithms are inherently sensitive to hyperparameter settings. Probabilistic Curriculum Learning (PCL) is a curriculum learning strategy designed to improve RL performance by structuring the agent's learning process, yet effective hyperparameter tuning remains challenging and computationally demanding. In this paper, we provide an empirical analysis of hyperparameter interactions and their effects on the performance of a PCL algorithm within standard RL tasks, including point-maze navigation and DC motor control. Using the AlgOS framework integrated with Optuna's Tree-Structured Parzen Estimator (TPE), we present strategies to refine hyperparameter search spaces, enhancing optimisation efficiency. Additionally, we introduce a novel SHAP-based interpretability approach tailored specifically for analysing hyperparameter impacts, offering clear insights into how individual hyperparameters and their interactions influence RL performance. Our work contributes practical guidelines and interpretability tools that significantly improve the effectiveness and computational feasibility of hyperparameter optimisation in reinforcement learning.

Their interactions are characterized by significant dynamism and heterogeneity. To address these challenges, we propose a MADCZ modeling approach. By constructing dynamic topological structures and spatiotemporal conflict zones, the model attains precise conflict identification and delivers interpretable decision support. First, a joint state space is established, defined as S = S A Vs S BVs S Peds S Road, (2) where S A Vs, S BVs, S Peds and S Road represent the state subspaces of A Vs, BVs, Peds, and road network, respectively. Each subspace is specifically defined as S V ehs = [ x, y,θ, v,l,c, p ] R 22 S Peds = [ x, y,θ, v,l, c ] R 10 S Road = nullnull G(V,E) | V R n 22, E { 0, 1} n nnull, (3) where x and y denote the horizontal and vertical coordinates of the traffic participants, θ [0, 360) is the heading angle, v represents the longitudinal velocity, l and c represent the lane position and traffic participant type, respectively, each encoded as a three-dimensional one-hot vector. G represents the road network topology, where each traffic participant is modeled as a node v i V, and E represents the connections among participants, representing sensor perception or vehicle-to-vehicle (V2V) communication relationships. Additionally, for vehicles, p denotes the relative motion information with respect to surrounding vehicles, defined as p = [ d j, v j], j = {f, r, lf, lr,rf, rr }, (4) where d j and v j denote the relative longitudinal distance and the relative velocity between vehicles, and f, r, lf, lr, rf, rr represent the neighboring vehicles at the front, rear, left front, left rear, right front, and right rear, respectively. If no neighboring vehicle is detected in a given direction, the relative longitudinal distance is assigned the maximum perception range and the relative velocity is set to zero.

-- This paper proposes a novel alternative to existing sim-to-real methods for training control policies with simulated experiences. Prior sim-to-real methods for legged robots mostly rely on the domain randomization approach, where a fixed finite set of simulation parameters is randomized during training. Instead, our method adds state-dependent perturbations to the input joint torque used for forward simulation during the training phase. These state-dependent perturbations are designed to simulate a broader range of reality gaps than those captured by randomizing a fixed set of simulation parameters. Experimental results show that our method enables humanoid locomotion policies that achieve greater robustness against complex reality gaps unseen in the training domain. Deep Reinforcement Learning (DRL) for robotic applications has gained significant attention due to its demonstrated robustness and versatility. Although DRL algorithms are capable of solving complex, high-dimensional control problems, commonly used on-policy methods often require a prohibitively large amount of data, posing a substantial challenge when collecting sufficient samples solely from real hardware. Moreover, the exploration process required for policy improvement in early training stages can raise safety concerns for both the physical robot and its operational environment.

Fleets of robo-taxis offering on-demand transportation services, commonly known as Autonomous Mobility-on-Demand (AMoD) systems, hold significant promise for societal benefits, such as reducing pollution, energy consumption, and urban congestion. However, orchestrating these systems at scale remains a critical challenge, with existing coordination algorithms often failing to exploit the systems' full potential. This work introduces a novel decision-making framework that unites mathematical modeling with data-driven techniques. In particular, we present the AMoD coordination problem through the lens of reinforcement learning and propose a graph network-based framework that exploits the main strengths of graph representation learning, reinforcement learning, and classical operations research tools. Extensive evaluations across diverse simulation fidelities and scenarios demonstrate the flexibility of our approach, achieving superior system performance, computational efficiency, and generalizability compared to prior methods. Finally, motivated by the need to democratize research efforts in this area, we release publicly available benchmarks, datasets, and simulators for network-level coordination alongside an open-source codebase designed to provide accessible simulation platforms and establish a standardized validation process for comparing methodologies. Code available at: https://github.com/StanfordASL/RL4AMOD

Hierarchical inductive biases are hypothesized to promote generalizable policies in reinforcement learning, as demonstrated by explicit hyperbolic latent representations and architectures. Therefore, a more flexible approach is to have these biases emerge naturally from the algorithm. We introduce Free Random Projection, an input mapping grounded in free probability theory that constructs random orthogonal matrices where hierarchical structure arises inherently. The free random projection integrates seamlessly into existing in-context reinforcement learning frameworks by encoding hierarchical organization within the input space without requiring explicit architectural modifications. Empirical results on multi-environment benchmarks show that free random projection consistently outperforms the standard random projection, leading to improvements in generalization. Furthermore, analyses within linearly solvable Markov decision processes and investigations of the spectrum of kernel random matrices reveal the theoretical underpinnings of free random projection's enhanced performance, highlighting its capacity for effective adaptation in hierarchically structured state spaces.

This paper introduces a novel reinforcement learning (RL) approach to scheduling mixed-criticality (MC) systems on processors with varying speeds. Building upon the foundation laid by [1], we extend their work to address the non-preemptive scheduling problem, which is known to be NP-hard. By modeling this scheduling challenge as a Markov Decision Process (MDP), we develop an RL agent capable of generating near-optimal schedules for real-time MC systems. Our RL-based scheduler prioritizes high-critical tasks while maintaining overall system performance. Through extensive experiments, we demonstrate the scalability and effectiveness of our approach. The RL scheduler significantly improves task completion rates, achieving around 80% overall and 85% for high-criticality tasks across 100,000 instances of synthetic data and real data under varying system conditions. Moreover, under stable conditions without degradation, the scheduler achieves 94% overall task completion and 93% for high-criticality tasks. These results highlight the potential of RL-based schedulers in real-time and safety-critical applications, offering substantial improvements in handling complex and dynamic scheduling scenarios.

Controlling instabilities in complex dynamical systems is challenging in scientific and engineering applications. Deep reinforcement learning (DRL) has seen promising results for applications in different scientific applications. The many-query nature of control tasks requires multiple interactions with real environments of the underlying physics. However, it is usually sparse to collect from the experiments or expensive to simulate for complex dynamics. Alternatively, controlling surrogate modeling could mitigate the computational cost issue. However, a fast and accurate learning-based model by offline training makes it very hard to get accurate pointwise dynamics when the dynamics are chaotic. To bridge this gap, the current work proposes a multi-fidelity reinforcement learning (MFRL) framework that leverages differentiable hybrid models for control tasks, where a physics-based hybrid model is corrected by limited high-fidelity data. We also proposed a spectrum-based reward function for RL learning. The effect of the proposed framework is demonstrated on two complex dynamics in physics. The statistics of the MFRL control result match that computed from many-query evaluations of the high-fidelity environments and outperform other SOTA baselines.

Improving decision-making capabilities in Autonomous Intelligent Vehicles (AIVs) has been a heated topic in recent years. Despite advancements, training machines to capture regions of interest for comprehensive scene understanding, like human perception and reasoning, remains a significant challenge. This study introduces a novel framework, Human Attention-based Explainable Guidance for Intelligent Vehicle Systems (AEGIS). AEGIS utilizes human attention, converted from eye-tracking, to guide reinforcement learning (RL) models to identify critical regions of interest for decision-making. AEGIS uses a pre-trained human attention model to guide RL models to identify critical regions of interest for decision-making. By collecting 1.2 million frames from 20 participants across six scenarios, AEGIS pre-trains a model to predict human attention patterns.

The Autonomy of Unmanned Aerial Vehicles (UAVs) in indoor environments poses significant challenges due to the lack of reliable GPS signals in enclosed spaces such as warehouses, factories, and indoor facilities. Micro Aerial Vehicles (MAVs) are preferred for navigating in these complex, GPS-denied scenarios because of their agility, low power consumption, and limited computational capabilities. In this paper, we propose a Reinforcement Learning based Deep-Proximal Policy Optimization (D-PPO) algorithm to enhance realtime navigation through improving the computation efficiency. The end-to-end network is trained in 3D realistic meta-environments created using the Unreal Engine. With these trained meta-weights, the MAV system underwent extensive experimental trials in real-world indoor environments. The results indicate that the proposed method reduces computational latency by 91\% during training period without significant degradation in performance. The algorithm was tested on a DJI Tello drone, yielding similar results.