Conceptcounterfactual post-training
Counterfactual Post-Training
A training method that actively modifies recorded action trajectories to generate future videos under diverse interaction configurations. It broadens the model's coverage of failure cases and invalid interactions, helping the model learn the consequences of actions that do not appear in the original demonstration data and improving robustness.
Timeline (last 90 days)
Oct 12
DreamTrue introduces counterfactual post-training, which generates future videos under a broader range of action and contact configurations by modifying recorded action trajectories, in order to broaden interaction coverage.
Unverified50%