Event
Uncertainty Projection Improves Cost Computation for Latent Space Planning
1 reports · 1 sources
Yoo Sung Jang et al. propose Latent Projection for Latent Planning to address the issue that, in latent-space world models, planning costs are disturbed by noise when the cost is taken as the ℓ2 distance between the predicted terminal state and the goal latent variable; they use the state dynamics covariance to restrict this cost to the k-dimensional subspace with the least uncertainty. This correction is a post-processing method that fits only one constrained matrix A and does not modify the encoder, predictor, rollout, or planner. The paper reports evaluation in six environments; in the harder planning settings of the 2D navigation tasks TwoRoom and PointMaze, the success rate improves by up to 54.7 percentage points compared with the ordinary ℓ2 cost; the paper has been accepted to the CoRL 2026 workshop on uncertainty modeling for robot world models.
Report timeline
- Latent Projection for Latent Planning: Improving World-Model Planning Costs with Latent-Space Projection
The paper proposes Latent Projection for Latent Planning, noting that planning costs in latent world models are often taken as the ℓ2 distance between the predicted terminal latent and the goal latent, while the latent itself contains noise whose source is precisely the regularizer used during training to prevent collapse.
arXiv cs.RO ·