Driving · 2026-10-21
Beckmann World Models: Direct Terminal-Map Learning for One-Step Action-Conditioned Prediction
Adam Lee, Shobhit Agarwal, June Pyo Suh
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
Beckmann World Models (BWM) adapt the terminal-map conservation idea to action-conditioned one-step video prediction. They learn a direct map from noise, visual history, and actions to future frames without a pretrained generative teacher or per-condition populations of generated futures. Endpoint-aligned training (source anchoring, motion/spatial supervision, and generated-history recovery) reduces errors and improves policy-score estimation under low-cost single-pass inference.
Why it matters
BWM targets the gap between terminal-map training on interpolated inputs and accurate prediction from the pure noise used at inference. The paper also emphasizes one-call-per-chunk deployment efficiency for downstream policy evaluation, reporting both prediction accuracy and policy-score estimation error with measured inference latency and training-step throughput/memory profiling.
Method
- Adapt Beckmann transport terminal-map learning to action-conditioned video prediction, learning a conditional terminal map Tθ(z, c) from noise and (history, actions) to future frames without a pretrained generative teacher.
- Use endpoint-aligned training: source anchoring directly supervises one-step prediction from initial noise; add motion/temporal differences, frozen VGG spatial features, and detached generated-history recovery for autoregressive use.
- Keep inference efficient: all components train the same map and the model retains one network call per future chunk; evaluate cost (calls/latency), policy evaluation accuracy, and training-step throughput/memory.
Limitation
The authors state that their claims about the learned map are empirical: “Merely imposing a pointwise identity on the target and conservation away from it does not prove uniqueness... Likewise, the population consistency argument... is not a convergence theorem for our adaptively weighted, source-regularized neural objective. Our claims about the learned map are empirical.”
Abstract (from arXiv)
Action-conditioned world models support planning and policy evaluation by predicting future states. However, learning an accurate one-step predictor can require substantial training computation. We introduce Beckmann World Models (BWM), which learn a direct map from noise, visual history, and actions to future frames without a pretrained generative teacher. We adapt the terminal-map conservation law of Beckmann Transport Models to action-conditioned video prediction, training directly from noise-data pairs without conditional populations of generated samples required by DriftWorld. However, transport consistency on interpolated inputs does not ensure accurate prediction from the pure noise used at inference. We address this gap with endpoint-aligned training that directly supervises the source prediction, emphasizes motion and spatial structure, and teaches recovery from generated histories. All components train the same map and retain one network call per future chunk. On PushT and Robomimic Can, the final model reduces 64-frame prediction MSE by 68% and 57%, respectively, relative to same-backbone regression baseline. For policy evaluation, BWM achieves the lowest mean IoU error among the evaluated models, 0.0100, at 24.8 ms per four-frame PushT prediction. On Robomimic Can, the BWM stage trained with motion supervision and generated histories achieves 3.3 times higher throughput and 78% lower peak memory than local one-step DriftWorld. In a latent-space extension to RT-1 and Bridge-V2, equal-update controls show that feature supervision under generated histories improves perceptual prediction. These findings show how terminal-map learning can combine efficient training with accurate one-step visual prediction.
Related papers
- De-Diffusing WAMs: Action-Sufficient Direct Prediction for Robot Control
- {\Delta}WAM: Distilling Action Tangent Fields into World Action Models
- MiniWAM: Learning Compact Future Targets for Efficient World-Action Modeling
- FutureWorlds: Learning Robotic World Models from Alternative Futures
- RAP-WAM: Learning World-Action Models through Predictive Rehearsal and Action Perturbations
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-16 · Indexed by Pub-AI: 2026-10-21 UTC
Beckmann World Models: Direct Terminal-Map Learning for One-Step Action-Conditioned Prediction
Preprint posted: 2026-10-16 · Original record ↗
- arxiv
- 2610.26903
Explicit source versions (1)
- v1 · 2026-10-16
- Preprint available
Source date: 2026-10-16 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (2)
- Official evidence ↗ · OAI arXiv created; not header datestamp (official metadata)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Action-conditioned world models support planning and policy evaluation by predicting future states. However, learning an accurate one-step predictor can require substantial training computation. We introduce Beckmann World Models (BWM), which learn a direct map from noise, visual history, and actions to future frames without a pretrained generative teacher. We adapt the terminal-map conservation law of Beckmann Transport Models to action-conditioned video prediction, training directly from noise-data pairs without conditional populations of generated samples required by DriftWorld. However, transport consistency on interpolated inputs does not ensure accurate prediction from the pure noise used at inference. We address this gap with endpoint-aligned training that directly supervises the source prediction, emphasizes motion and spatial structure, and teaches recovery from generated histories. All components train the same map and retain one network call per future chunk. On PushT and Robomimic Can, the final model reduces 64-frame prediction MSE by 68% and 57%, respectively, relative to same-backbone regression baseline. For policy evaluation, BWM achieves the lowest mean IoU error among the evaluated models, 0.0100, at 24.8 ms per four-frame PushT prediction. On Robomimic Can, the BWM stage trained with motion supervision and generated histories achieves 3.3 times higher throughput and 78% lower peak memory than local one-step DriftWorld. In a latent-space extension to RT-1 and Bridge-V2, equal-update controls show that feature supervision under generated histories improves perceptual prediction. These findings show how terminal-map learning can combine efficient training with accurate one-step visual prediction.