A robot that learns by trial and error inside its own learned simulator, called a world model, can avoid most real-world crashes by testing risky plans in imagination first. That protection is only as good as the model. When the robot's camera produces artifacts its training simulator never showed, such as motion blur, vibration, glare, or dropped frames, the model can mistake a corrupted view of a wall for empty space and plan straight into it. It can also treat a harmless flicker as a threat and freeze. We propose to split the world model's internal state into separate parts. The safety predictor may read only the parts that the cost signal shows are relevant to danger. A further part, which has no memory from one step to the next, absorbs per-frame sensor corruption so that it cannot leak into the model's predictions about the future.
The test of the idea is concrete. Take vision-based constrained navigation and driving tasks. Corrupt the camera at test time with noise families that were never seen in training, including noise whose strength depends on the agent's own actions. A safe agent that plans with this factored model should stay near its safety budget, while an otherwise identical planner with an undivided latent state should either exceed the budget, because it believes unsafe paths are safe, or collapse into inactivity.
Core Claim
Model-based safe exploration currently plans inside an undivided latent state. SafeDreamer adds Lagrangian cost constraints to the DreamerV3 recurrent world model and plans over imagined rollouts. A Lagrangian constraint folds the safety cost into the objective with a weight that is adjusted automatically during training. SafeDreamer reports near-zero cost on Safety-Gymnasium and MetaDrive. It was never tested under sensor corruption, and its own authors flag "hallucinated safety" when the model mispredicts hazards.
Separately, factored world models keep only the variables an agent needs and discard background noise. IFactor splits the latent state into four blocks, defined by whether actions affect a block and whether reward depends on it. Exogenous-state methods such as Predictive Path Elimination filter out background processes that actions cannot influence. None of these methods defines relevance by safety cost. None distinguishes a persistent background process, like a moving pedestrian, from memoryless observation corruption, like a blurred frame.
Our claim is that the right factorization for safe exploration under sensor noise has three properties. Relevance is defined by the cost signal, not only by reward. Uncontrollable but cost-relevant blocks, such as moving hazards, are kept explicitly. Sensor corruption is assigned to a separate block whose prior has no temporal transition. With $s_t$ split into a controllable cost-relevant block $s^{ac}_t$, an uncontrollable cost-relevant block $s^{\bar{a}c}_t$, a nuisance block $s^{n}_t$, and a per-step sensor block $\eta_t$, the model is
$$p(o_t \mid s^{ac}_t, s^{\bar a c}_t, s^n_t, \eta_t),\quad s^{ac}_{t+1}\sim f(s^{ac}_t, s^{\bar a c}_t, a_t),\quad s^{\bar a c}_{t+1}\sim g(s^{\bar a c}_t),\quad \eta_t \sim q(\eta \mid a_t)\ \text{independent across } t,$$
$$\hat c_t = h(s^{ac}_t, s^{\bar a c}_t).$$
In these equations, $o_t$ is the camera observation, $a_t$ the action, and $\hat c_t$ the predicted safety cost. Reward-relevant blocks are handled as in IFactor and omitted here for brevity.
Two design choices carry the claim. First, the sensor block $\eta_t$ may depend on the current action. Motion blur scales with turning rate. It may not persist into the next step. Without this block, an inverse-dynamics objective sees blur that is predictable from the action and labels it controllable. The section on factored models under sensor noise in the research file identifies this failure mode, and it would put the corruption straight into the block the safety predictor reads. Second, the cost head $h$ reads only the two cost-relevant blocks. Variation in the nuisance block and sensor block therefore cannot change the predicted cost, which is exactly what protects the agent against both false safety and false alarms.
Research Program
We would build the model on DreamerV3, keeping SafeDreamer's planners unchanged. These are OSRP-Lag, which plans online with a constrained cross-entropy search under a PID-controlled Lagrange multiplier, and BSRP-Lag, which trains a background actor with an augmented Lagrangian. Keeping the planners fixed means any difference comes from the representation.
The block split is enforced with mutual-information terms in the style of IFactor. These terms push the controllable blocks to be predictable from actions and push the cost-relevant blocks to be predictive of current and future cost. Two additions are specific to this proposal. The first is a temporal-independence penalty on $\eta_t$: the model is penalized when $\eta_t$ helps predict $s_{t+1}$ or $\eta_{t+1}$. The second is a cost-invariance check: the model is trained so that resampling $\eta_t$ or $s^n_t$ leaves $\hat c_t$ unchanged.
On the analysis side, we would ask whether IFactor's block-wise identifiability argument extends when the observation includes an action-dependent but memoryless corruption. The intuition is that temporal persistence identifies what belongs in the world-state blocks. We would state the conditions under which this holds, and also where it fails. For example, it fails when sensor drift is slow, as with thermal drift in an inertial sensor, and therefore looks persistent.
Baselines
The naive reference point is SafeDreamer (OSRP-Lag and BSRP-Lag) trained on clean simulator observations and tested on corrupted ones. It shows how much sensor noise actually breaks a model-based safe agent. The strongest existing method is the same SafeDreamer trained with observation corruption from the Robust Gymnasium observation disruptors at training time. It is the strongest competitor because it attacks the identical problem with no architectural change.
On clean vision tasks, SafeDreamer reports a 94.3% cost reduction relative to LAMBDA, the earlier model-based safe RL method. On MetaDrive it reports a 100% arrival rate with an average episode cost of 2.2, against 14.1 for PPO-based reward shaping (SafeDreamer). No reported result was found on Lacuna for SafeDreamer under observation disruption. That number is the first measurement this project would produce.
A model-free reference is PPO-Lag with domain randomization. On Safety-Gymnasium velocity tasks, PPO-Lag reduced safety violations by 98% relative to plain PPO, at the price of a 45% drop in total reward (Safety Gymnasium).
Three ablations isolate the proposal's ingredients. The first is IFactor's reward-relevance split dropped into SafeDreamer, which is the obvious combination of the two lines of work. The second is our model without the memoryless sensor block. The third is our model with the cost head allowed to read all blocks.
Every method gets the same environment steps, the same world-model size, the same planner settings, and an equal hyperparameter-search budget. Every method trained with augmentation sees the same training noise families. Beating the strongest baseline means two things on held-out noise families. Episode cost must stay within the budget where augmented SafeDreamer exceeds it, and return must not be lower. Matching augmented SafeDreamer on the noise seen in training is the expected outcome and does not count as a win.
Why not just…?
Why not just train SafeDreamer with heavy noise augmentation? This is the real threat. If the deployment noise is known in advance and simulated faithfully, augmentation may match our method. The sim-to-real problem is precisely that deployment noise is not known in advance. The experiments therefore separate noise families seen in training from held-out ones. If augmented SafeDreamer matches us on held-out families, the factorization adds nothing, and we would report that. Our method can also be combined with augmentation, so we would test the two together as well.
Why not just use IFactor's existing reward-relevant split? Reward and cost can come apart. A hazard that never affects reward but does incur cost would be classified by a reward-based split as irrelevant noise and discarded. That is the hallucinated-safety failure in a new form. The reward-relevance ablation tests this directly. We would measure, with linear probes, how well hazard positions can be recovered from each model's relevance blocks, and compare episode costs. IFactor reports an identifiability score above 0.9, but for reward-relevant factors only (IFactor).
Why not just add uncertainty-triggered shielding to the undivided model? An uncertainty estimate can be fooled by the same corruption it is meant to detect. In CARLA driving, Monte Carlo dropout's safety score fell to 0.00 in rain, while Hamiltonian Monte Carlo inference achieved 90% safety (Uncertainty Quantification with Statistical Guarantees). We would include an ensemble-disagreement version of SafeDreamer that falls back to a conservative action when the ensemble disagrees, and report how often it intervenes alongside its cost. A shield that keeps the agent safe only by stopping it is the "safe but useless" outcome. A model-free reachability shield as in Sim-to-Lab-to-Real reads raw RGB frames and would be expected to share this fragility.
Why not just use a classical perception front end? Deep Drone Acrobatics shows that geometric feature tracks transfer from simulation to real flight: raw-pixel models had 0% success on new backgrounds while the feature-track model had 100%. That approach depends on feature tracking working at all, and it fails in dark or textureless scenes. It is a different operating regime, and we would not run it as a baseline.
Proposed Experiments
The testbed is the vision-based Safety-Gymnasium navigation tasks with static and moving hazards, together with MetaDrive. Observations are wrapped with Robust Gymnasium-style observation disruptors, and we add distractor backgrounds of the kind used in IFactor's evaluations.
Training uses i.i.d. Gaussian pixel noise and random pixel dropout. Held-out test noise comes in three kinds. The first is action-correlated: motion blur proportional to turning rate, and frame vibration proportional to throttle. The second is temporal: random frame latency and occasional frozen frames. The third is structured artifacts such as glare patches and lens occlusion. Robust Gymnasium found that how often a disruption occurs hurts performance more than its size, so we sweep both frequency and magnitude.
We measure the following. Episode cost and normalized return are measured at deployment. Cumulative cost is measured during training, because safety during exploration is the point. We also measure a hallucinated-safety rate: the fraction of planner-approved trajectories that actually incurred cost. A false-alarm rate measures how often the agent stalls in states where no cost was possible. Probe accuracy for hazard state from $s^{ac}$ and $s^{\bar a c}$ shows whether hazard information is kept, and probe accuracy for the corruption level from those same blocks shows whether noise leaks in.
The decisive checks are four. First, on action-correlated held-out noise, removing the memoryless sensor block should make the corruption level recoverable from $s^{ac}$ and raise both the hallucinated-safety rate and episode cost. If it does not, the action-correlated sensor channel is unnecessary. Second, moving hazards should be recoverable from $s^{\bar a c}$ under the cost-relevant split and lost under the reward-relevant ablation. Third, on held-out noise, our planner should stay within the cost budget where augmented SafeDreamer exceeds it, at equal return. A tie would mean the idea reduces to data augmentation. Fourth, as a stress test of the identifiability argument, slow drift noise should defeat the memoryless block, which would confirm the stated boundary of the method rather than hide it.
References
- SafeDreamer: Safe Reinforcement Learning with World Models
- Learning World Models with Identifiable Factorization
- Provably Filtering Exogenous Distractors using Multistep Inverse Dynamics
- Robust Gymnasium: A Unified Modular Benchmark for Robust Reinforcement Learning
- Safety Gymnasium: A Unified Safe Reinforcement Learning Benchmark
- Uncertainty Quantification with Statistical Guarantees in End-to-End Autonomous Driving Control
- Sim-to-Lab-to-Real: Safe RL with Shielding and Generalization Guarantees
- Deep Drone Acrobatics (Extended Abstract)