Robot world models learn to notice small action changes: policy-success correlation on DROID rises from 0.47 to 0.79
Synopsis
Robot world models can generate realistic manipulation video yet often ignore small action changes, predicting the same outcome for actions with different real consequences. AFFECT injects logged perturbations into training actions and requires generated videos to respond to those perturbations as recorded videos do, adding action-sensitivity supervision without architectural changes or repeated executions from the same initial state. On the DROID real-robot configuration, the Pearson correlation between policy success rates inside the DreamDojo world model and on the real robot rises from 0.47 to 0.79, with FVD lower by up to 23%, and planning with the trained model improves task success.
Interpretation
Robot world models begin to distinguish small differences in actions instead of only producing plausible-looking video. World models trained only to reconstruct recorded video tend to predict similar outcomes for actions with different real consequences, which misleads policy evaluation and planning. On the DROID real-robot configuration, the Pearson correlation between policy success rates in the DreamDojo world model and on the real robot rises from 0.47 to 0.79, alongside FVD reductions of up to 23%.
The approach perturbs actions during data collection and teaches the model the visual changes those perturbations cause. Training data come from a policy trained on task-agnostic human teleoperation, with temporally correlated noise injected into each action chunk and logged, so no repeated executions from the same initial state and no outcome labels are needed. Perturbation and video features are extracted by frozen encoders, and the loss compares their covariance in recorded and generated videos; across UR5e simulation and the DROID and ALOHA-2 real-robot configurations, correlation improves in all six backbone-embodiment combinations.
With better action sensitivity, the world model also changes how it behaves in closed-loop evaluation and candidate-action selection. For planning, an ensemble of policy checkpoints proposes candidate actions, the world model imagines each candidate's future video, and a fixed success estimator scores it, so only the world model changes. On five OGBench pick-and-place tasks with UR5e, average success rises from the random-selection baseline to DreamDojo and further to DreamDojo with AFFECT; on the ALOHA-2 long-horizon task the average gain is about several percentage points.
Perspective
The demonstrated scope is the UR5e simulation environment and two real-robot setups, a Franka Research 3 in the DROID configuration and a bimanual ALOHA-2, with training data collected under the perturbation-play protocol and evaluation in closed-loop policy rollouts and candidate-action selection.
The loss constrains only responses the chosen feature bank can detect, leaving responses outside that bank unconstrained; covariances are estimated from finitely many windows, so finite-sample noise remains; on ALOHA-2, DreamDojo's FVD rises slightly from 183 to 188; and with 20% of the training data the gain over the baseline is smaller.
