Disentangle the contribution of future information from supervision and model capacity

Determine how to separate the contribution of explicit future occupancy information from the effects of additional network capacity, supervision, future tokens, and the analytic risk vector in the LOOP occupancy-forecasting policy.

Background

The paper’s rollout ablation removes several components simultaneously: the recurrent rollout, prediction supervision, future occupancy tokens, and the risk vector. Consequently, the reported performance difference cannot isolate whether any benefit comes specifically from explicit future information or instead from increased capacity, auxiliary supervision, or additional control features. The authors identify disentangling these effects as an unresolved problem for interpreting the causal value of forecasting.

References

At this operating point the benefit comes from conditioning on map-space futures at all, not from their accuracy -- which is also why zeroing the channel, an implausible empty scene, is what destroys performance (Sec.~\ref{sec:res_audit}). Second, the rollout ablation removes network, supervision, tokens and risk vector together ($0.30$\,M of $0.94$\,M parameters), so separating explicit future information from supervision and capacity remains open, as does a tracking-plus-velocity-obstacle pipeline on the same LiDAR input.

— Predict Before You Step: Auditable Occupancy Forecasting for Dynamic Obstacle Avoidance under Sparse Guidance  (2609.25969 - Mao et al., 22 Sep 2026) in Section Discussion and Limitations