Latent world models are a promising approach for learning state representations and dynamics directly from high-dimensional observations, enabling robot control in hard-to-model settings. However, control performance ultimately depends on the latent representation encoding the required information for the task. In this work, we study latent-space safe control problems and show how partial observability can induce control failures when safety-relevant information is not preserved in the latent state. Specifically, we identify two world model failure modes: estimation gaps, where current observations do not reveal safety-critical quantities (e.g., temperature in a cooking task), and prediction gaps, where failures are observable once they occur but cannot be reliably anticipated from available observations. We introduce two diagnostics for these gaps: a mutual-information-based measure of safety observability and a rollout-based measure of future safety predictability. Finally, we present mitigation strategies for each failure mode: privileged multimodal supervision for estimation gaps and conformal risk calibration for prediction gaps. Across two hardware case studies -- using unimodal RGB world models and multimodal RGB+Tactile and RGB+Thermal variants -- we show that these mitigation strategies improve the safety of a Franka Research 3 manipulator on challenging cooking tasks under partial observability, albeit with increased conservativeness. More broadly, our work raises the question of when world model state representations are sufficient for reliable robot control
Figures & tables
Figure 1: How well do latent world models (WMs) understand safety constraints? We find that WMs can suffer from estimation gaps (left) where the latent state fails to encode safety-relevant states that are hard to observe (e.g., temperature from RGB images) and prediction gaps (right) where failures are identifiable from the latent state once they happen (e.g., a spill) but cannot be reliably predicted with current observations because of unobservables (e.g., if the bottle is filled).
Wax Melting
Rice Pouring
Metric
RGB
IR
MM
RGB
Tactile
MM
H(Y)I(Y;X)
0.221
0.801
0.853
0.962
0.221
0.968
Acc.
0.870
0.969
0.979
0.994
0.838
0.995
B. Acc.
0.626
0.983
0.988
0.988
0.687
0.988
Table 1: Safety-Relevant Signal in Observations. Mutual information estimates and classification accuracies.
Table 2: Safety Outcome Prediction. Outcomes of 16-step open-loop trajectories. The right table reports results only on samples where RGB and MM disagree on safety outcomes; rows follow the same method ordering as the left table.
Figure 2: Real vs. Imagined Observations & Safety Labels. Open-loop predictions of a held-out subtrajectory reveal that WMRGB misjudges whether lifting the wax plate will cause failure. In contrast, observing temperature via IR images allows the WMMM to correctly classify that this action leads to a safe outcome.
World Models & Respective Safety Filters
Metric
WMRGB
WMMM
WMRGB-MM
Lift Rate [95% CI]
0.15[0.04,0.39]
1.0[0.80,1.0]
1.0[0.80,1.0]
Pixel Intensity at Intervention
0.587±0.303
0.695±0.073
0.522±0.208
Table 3: Hardware: Closed-Loop Safety Filter Performance. Lift rates over 20 trials with 95% Wilson-CC confidence intervals [ 32 ] . WMRGB fails to lift reliably. WMMM intervenes before failure, while WMRGB-MM intervenes more variably, likely because it cannot directly observe safety-relevant quantities.
Figure 3: (Left) Ground-truth pixel intensities when each safety filter intervenes. The WMMM policy that takes in IR images concentrates just prior to the threshold used for failure labeling, whereas WMRGB and WMRGB-MM use only RGB as input are spread more conservatively, indicating they are unable to discern when to filter. (Right) Distribution of predicted failure probabilities during conformal calibration. The WMRGB frequently fails to predict spills leading to conservative thresholds.
Bottle
Model
Pour Motion
Fallback
[95% CI]
[95% CI]
Filled
RGB
0.14 [0.06, 0.27]
0.86 [0.73, 0.94]
MM
0.14 [0.06, 0.27]
0.86 [0.73, 0.94]
Empty
RGB
0.18 [0.09, 0.32]
0.82 [0.68, 0.91]
MM
0.60 [0.45, 0.73]
0.40 [0.27, 0.55]
Table 4: Results: MPS for 50 trials. Rates are reported with 95% Wilson-CC confidence intervals [ 32 ] . After conformal calibration to control false acceptance, the multimodal MPS controller retains greater control authority in the empty-bottle setting while maintaining fallback behavior in higher-risk settings.
Figure 4: Experiment Testbeds. In simulation (left) and Wax Melting (right), we collect paired RGB and infrared (IR) observations. In both settings, the safety-relevant state variable is heat, which is directly observable from IR but only partially observable from RGB.
Figure 5: Safety Controllers without & with Observability. In the Thermal Unicycle example, the robot only appears burnt in RGB after it leaves the hot plate. Left: The RGB-only safety policy is myopic: it stays in the hot region, avoiding seeing failure. Right: When the WM is trained with RGB+IR observations, the resulting safety controller accelerates out of the hot region to prevent overheating.
Modality
I(Y;X)/H(Y)↑
Acc. ↑
B. Acc. ↑
RGB
0.182
0.864
0.738
IR
0.909
0.982
0.966
RGB+IR
0.903
0.981
0.976
Table 5: Simulation Observation Signal Metrics. Observation-level safety signal metrics for the simulated Thermal Unicycle task.
Eval.
Method
TS ↑
FS ↓
FU ↓
TU ↑
F1↑
Latent State
RGB
0.623
0.013
0.104
0.272
0.914
RGB+IR
0.701
0.002
0.026
0.283
0.981
Open-loop Pred.
RGB
0.643
0.012
0.089
0.260
0.927
RGB+IR
0.693
0.012
0.035
0.264
0.967
Table 6: Simulation Safety Classification Metrics. Latent-state and 16-step open-loop prediction metrics for the simulated Thermal Unicycle task.
Hyperparameter
Thermal Unicycle
Wax Melting
RGB Image Dimension
[128, 128, 3]
[224, 224, 3]
IR Image Dimension
[128, 128, 1]
[224, 224, 1]
Action Dimension
2
7
Stochastic Latent
Gaussian
Gaussian
Latent Dim (Deterministic)
512
512
Latent Dim (Stochastic)
32
32
Table 7: Dreamer RSSM hyperparameters for the Thermal Unicycle and Wax Melting experiments.
Hyperparameter
Value
Actor Architecture
[512, 512, 512]
Critic Architecture
[512, 512, 512]
Normalization
LayerNorm
Activation
ReLU
Discount Factor γ
0.9999
Learning Rate (Critic)
1e-3
Table 8: DDPG hyperparameters for latent HJ safety filter synthesis.
World models offer a powerful substrate for safety reasoning in high-dimensional robotic systems, but they are also fallible: their predictions can be biased, miscalibrated, or confidently wrong. This creates a central challenge for latent-space safety filters, which often learn Hamilton-Jacobi safety value functions on the dynamics of a world model. If the world model is incorrect, the resulting value function can inherit its errors and produce overconfident safety estimates. Existing latent safety filters often rely on auxiliary signals such as ensemble disagreement or value-target consistency residuals for adaptation, but these signals can remain small even when the world model's predictions deviate from observations. We propose an adaptive latent safety filter that calibrates safety reasoning using directly observed world-model error. Our method uses Adaptive Conformal Inference to construct online uncertainty sets from discrepancies between predicted and observation-inferred latent states, then evaluates safety pessimistically by minimizing the learned value function over these sets. This allows the filter to remain minimally conservative when the world model is accurate, while becoming more cautious when observations reveal model mismatch. We provide a finite-time coverage guarantee for the adaptive uncertainty radius. Through simulation and hardware experiments, we show that our method significantly reduces failures relative to state-of-the-art latent safety filters while preserving task completion.
John Cao, Somil Bansal
Department of Aeronautics and Astronautics, Stanford University
In this paper, we study robust decision-making in the latent space of world models (WMs). Robust optimization is a mathematical framework where, given explicitly specified dynamics and physically meaningful disturbances, a robot can select actions that remain effective even under worst-case disturbances. However, applying this principle to the learned latent space of WMs introduces a fundamental challenge: because WMs have fully learned state spaces and dynamics inferred from high-dimensional observations, it is unclear how to define latent-space disturbances that faithfully represent uncertainty in the underlying system. Our key idea is to model a latent-space disturbance as a perturbation to the learned latent dynamics that induces pessimistic but plausible transitions. Specifically, we construct a set of plausible latent dynamics by combining a dynamics-aware similarity metric that captures plausible transitions with out-of-distribution detection that excludes implausible latent states. We calibrate this uncertainty set over latent dynamics using conformal prediction, ensuring that WM imaginations induced by the latent disturbance remain plausible without becoming overly pessimistic. We then jointly optimize robust robot actions and the worst-case latent disturbances through game-theoretic optimization. We leverage this latent-space robust optimization to robustify policy steering, considering two paradigms: latent safety filtering and sample-and-verify steering of a generative control policy. Our controlled simulation experiments show that our latent disturbance enables robust decision-making directly in WM latent spaces, and hardware experiments with a Franka manipulator show that modeling latent disturbances enables robust policy steering, reducing failures by 70% in safety filtering and 54% in sampling-based policy steering. Project website: https://junwon.me/LatentDisturbance/.
Junwon Seo, Andrea Bajcsy
The authors are with the Robotics Institute, Carnegie Mellon University, Pittsburgh, PA, USA.
We present SLS^2, a framework for safe feedback motion planning from pixels using robust model predictive control (MPC) in learned latent world models. Our approach trains an action-conditioned joint-embedding world model with compact Markovian latent states, enabling efficient gradient-based trajectory optimization through learned latent dynamics. To enforce safety for the true system despite imperfect latent predictions, we inform a GPU-accelerated system level synthesis (SLS) robust MPC scheme with conformal prediction to obtain calibrated latent error bounds and robust latent-space constraint sets. We further learn and conformalize a latent constraint checker, allowing the SLS planner to impose probabilistic safety constraints during closed-loop execution. We evaluate our method on vision-based control tasks, where it improves both goal-reaching performance and safety over latent world-model and safe-planning baselines.