Latent Dynamics · 2026-10-08
Directed Temporal Representations for Offline Visual Control
Chenyang Yuan, Haoyu Wang, Zhuo Sun, Xiaoyuan Cheng
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
DTRC learns a directed temporal quasimetric for offline visual goal-conditioned control on top of frozen LeWorldModel features. It calibrates local distance with short temporal offsets, bootstraps long-horizon reachability, and enforces action-conditioned consistency. The resulting distance change defines goal-relative temporal progress, used as a temporal critic for direct goal-conditioned policy learning (no test-time trajectory search).
Why it matters
For control, the authors argue latent geometry must match temporal reachability and temporal progress (including directionality and local scale), not just predictive similarity. DTRC provides an offline temporal-supervision objective that yields goal-relative temporal advantage for training a direct policy, and reports empirical gains versus planning and direct-policy baselines.
Method
- Learn control representation on frozen LeWM features via a directed temporal quasimetric (symmetric metric + order-sensitive residual).
- Use offline trajectory supervision: short-range temporal offsets calibrate distance scale; bootstrapped targets propagate temporal reachability; action-conditioned consistency aligns with local transitions.
- Convert distance reduction across transitions into goal-relative temporal progress/advantage to weight recorded actions (and evaluate policy alternatives); train a direct goal-conditioned flow-policy (no iterative trajectory search).
Abstract (from arXiv)
Predictive world models provide compact visual representations for control. Control requires a latent geometry aligned with temporal reachability rather than predictive similarity alone. We introduce Directed Temporal Representations for Control (DTRC), which learns such a geometry from offline visual trajectories on top of frozen LeWorldModel (LeWM) features. DTRC constructs a directed temporal quasimetric over the learned control representation. Short-range temporal offsets calibrate the distance scale. Bootstrapped targets extend temporal reachability across longer horizons. Action-conditioned consistency aligns the representation with local transition dynamics. The resulting distance estimates temporal reaching cost, and its change across a transition defines goal-relative temporal progress. We use this progress signal as a temporal critic for direct goal-conditioned policy learning. Model-assisted targets provide an additional training-time refinement under behavior-support and dynamics-agreement constraints. Across ten visual control tasks, DTRC achieves strong goal-conditioned control performance relative to planning and direct-policy baselines. Held-out diagnostics on the four LeWM tasks show consistent short-range temporal calibration, task-dependent long-range and directional structure, and positive transition-level progress. Temporal supervision improves the same flow-policy parameterization across all four LeWM tasks, while the resulting policy acts directly without iterative trajectory search at test time.
Related papers
- Beyond Action Prediction: Learning Search-Free Planning in Latent Space
- Pretrained Features as Teachers, Not States: Compact Latent World Models for Visual Planning
- Beyond a single latent space: a dual-latent world model for long-horizon planning
- TS-WAM: Temporal Latent Actions and World-Action Coupling for Vision-Language Robot Control
- Video2STL: Grounding VLM-Generated Temporal Specifications for Robot Learning
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-06 · Indexed by Pub-AI: 2026-10-08 UTC
Directed Temporal Representations for Offline Visual Control
Preprint posted: 2026-10-06 · Original record ↗
- arxiv
- 2610.08960
Explicit source versions (1)
- v3 · 2026-10-12
- Explicit source version detected · v3
Source date: 2026-10-12 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:id explicit version (official metadata)
- Preprint available
Source date: 2026-10-06 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Predictive world models provide compact visual representations for control. Control requires a latent geometry aligned with temporal reachability rather than predictive similarity alone. We introduce Directed Temporal Representations for Control (DTRC), which learns such a geometry from offline visual trajectories on top of frozen LeWorldModel (LeWM) features. DTRC constructs a directed temporal quasimetric over the learned control representation. Short-range temporal offsets calibrate the distance scale. Bootstrapped targets extend temporal reachability across longer horizons. Action-conditioned consistency aligns the representation with local transition dynamics. The resulting distance estimates temporal reaching cost, and its change across a transition defines goal-relative temporal progress. We use this progress signal as a temporal critic for direct goal-conditioned policy learning. Model-assisted targets provide an additional training-time refinement under behavior-support and dynamics-agreement constraints. Across ten visual control tasks, DTRC achieves strong goal-conditioned control performance relative to planning and direct-policy baselines. Held-out diagnostics on the four LeWM tasks show consistent short-range temporal calibration, task-dependent long-range and directional structure, and positive transition-level progress. Temporal supervision improves the same flow-policy parameterization across all four LeWM tasks, while the resulting policy acts directly without iterative trajectory search at test time.