Felipe Andrade G. Tommaselli

Francisco Affonso, Felipe Andrade G. Tommaselli, Juliano Negri et al.

Traditional RL-based locomotion controllers often suffer from low data efficiency, requiring extensive interaction to achieve robust performance. We present a model-based reinforcement learning (MBRL) framework that improves sample efficiency for quadrupedal locomotion by appending synthetic data to the end of standard rollouts in PPO-based controllers, following the Dyna-Style paradigm. A predictive model, trained alongside the policy, generates short-horizon synthetic transitions that are gradually integrated using a scheduling strategy based on the policy update iterations. Through an ablation study, we identified a strong correlation between sample efficiency and rollout length, which guided the design of our experiments. We validated our approach in simulation on the Unitree Go1 robot and showed that replacing part of the simulated steps with synthetic ones not only mimics extended rollouts but also improves policy return and reduces variance. Finally, we demonstrate that this improvement transfers to the ability to track a wide range of locomotion commands using fewer simulated steps.

Marco S. Tayar, Lucas K. de Oliveira, Felipe Andrade G. Tommaselli et al.

Autonomous UAV inspection of confined industrial infrastructure, such as ventilation ducts, demands robust navigation policies where collisions are unacceptable. While Deep Reinforcement Learning (DRL) offers a powerful paradigm for developing such policies, it presents a critical trade-off between on-policy and off-policy algorithms. Off-policy methods promise high sample efficiency, a vital trait for minimizing costly and unsafe real-world fine-tuning. In contrast, on-policy methods often exhibit greater training stability, which is essential for reliable convergence in hazard-dense environments. This paper directly investigates this trade-off by comparing a leading on-policy algorithm, Proximal Policy Optimization (PPO), against an off-policy counterpart, Soft Actor-Critic (SAC), for precision flight in procedurally generated ducts within a high-fidelity simulator. Our results show that PPO consistently learned a stable, collision-free policy that completed the entire course. In contrast, SAC failed to find a complete solution, converging to a suboptimal policy that navigated only the initial segments before failure. This work provides evidence that for high-precision, safety-critical navigation tasks, the reliable convergence of a well-established on-policy method can be more decisive than the nominal sample efficiency of an off-policy algorithm.