Robotics Research Hub
Generalization of World Models under Environmental Variability for Vision-based Quadrotor Navigation

Generalization of World Models under Environmental Variability for Vision-based Quadrotor Navigation

Featured

Luca Zanatta, Grzegorz Malczyk, Kostas Alexis · N/A · 2026

sim-to-real reinforcement-learning navigation vision world-models
PDF arXiv

Framework

N/A

License

N/A

Stars

N/A

Summary

World models, learned generative models that predict how an environment evolves, have become a promising tool for sample-efficient robot learning. Yet how robust they are to environmental variability remains poorly understood.

Abstract Summary

World models, learned generative models that predict how an environment evolves, have become a promising tool for sample-efficient robot learning. Yet how robust they are to environmental variability remains poorly understood. To address this, we conduct a systematic study using vision-based quadrotor navigation as a testbed problem, training DreamerV3-based world models under varying levels of environmental randomness and evaluating them across all levels through cross-environment validation, spanning both Self-Supervised Learning (SSL) pretraining and Reinforcement Learning (RL) fine-tuning. We then deploy all world models and associated navigation policies on a real quadrotor in unseen environments, including an open-loop run where the model receives just 2.5s of real sensory input before all sensors are cut off, leaving the system to navigate entirely in imagination over a 12m traverse. Our results show that world model robustness during SSL pretraining is a strong predictor of sim-to-real transfer: every model that generalized well in cross-environment SSL validation deployed successfully in the real world, passing through gaps as narrow as 0.67m, whereas the model that dominated simulation policy evaluation failed on the real platform. We further identify (a) the discrete latent size and (b) the training-sequence length as the dominant factors governing world model quality.

Key Points

  • Uses simulation or synthetic data for training
  • Builds predictive world models for planning
  • Focuses on generalization across environments
  • Incorporates tactile or force feedback for robust interaction

Problem Formulation

This work investigates the limits and potential of mbrl through predictive world models for autonomous collision-free 3D navigation in unknown environments using onboard depth perception. Let us denote the inertial frame as $\mathcal{I}$, the body-fixed frame as $\mathcal{B}$.

Problem Formulation — Generalization of World Models under Environmental Variability for Vision-based Quadrotor Navigation
Figure 2: Method overview. First, the data is collected per environmental randomness level. Next, the DreamerV3 world model is trained by self-supervised reconstruction. Finally, an actor and critic are trained in imagination with world-model fine-tuning: after a real-observation context (green), the model rolls out in pure imagination (orange), hallucinating future observations.

Method

The method comprises three stages, illustrated in Figure 2 : (a) data collection under varying levels of environmental randomness ( Section 3.1 ), (b) ssl pretraining of DreamerV3 world models with hyperparameter optimization ( Section 3.2 ), and (c) rl fine-tuning of the best world model per randomness level ( Section 3.3 ). All stages are conducted in the AerialGym simulator [ 11 ] .

Method — Generalization of World Models under Environmental Variability for Vision-based Quadrotor Navigation
Figure 3: Hyperparameter sweep results: each cell shows all runs per environment (grey) with the best configuration highlighted. Stages S1–S4 sweep rssm size, batch length, discrete latent size, and seed, respectively, in sequence. For more details, please refer to Appendix B .

Results

Our evaluation follows three stages. We first examine self-supervised pretraining, analyzing hyperparameter sensitivity and cross-environment reconstruction quality in Section 4.1 .

We then evaluate rl fine-tuning in terms of policy performance and imagination fidelity in Section 4.2 , concluding by the deployment of all models on a real quadrotor in unseen environments, in both closed-loop and open-loop imagination in Section 4.3 . The sequential sweep is shown in Figure 4 .

Results — Generalization of World Models under Environmental Variability for Vision-based Quadrotor Navigation
Figure 4: Hyperparameter sensitivity across environments. The relative loss gap $\Delta$ between the worst and best value of each hyperparameter, per environment ($\mathrm{L}1$-$\mathrm{L}4$).

Sources and Demonstrations

Figures are reproduced from Zanatta et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.

Share

Related Papers

A Sim-to-Real Integration Pipeline for Training and Deployment of Chunk-Based VLA Manipulation Policies
arXiv preprint

A Sim-to-Real Integration Pipeline for Training and Deployment of Chunk-Based VLA Manipulation Policies

Mathilde Kappel, Clémence Grislain, Mohamed Chetouani et al.

Vision-Language-Action (VLA) models have become a prominent paradigm for mapping multimodal inputs, including semantic instructions, visual observations of the scene, and proprioceptive observations, to robot actions.

manipulation vision vla reinforcement-learning sim-to-real control learning-from-demonstration benchmark
PDF Advanced