Freeze FlowMo ICRA 2027 manuscript and clarify evaluation definitions

Clarify velocity-probe targets, task-specific station-keeping success and stopping, and planning initialization. Add temporal and sparse-memory related work. Preserve all experiments, numerical results, weights, data and figure sources. Update the eight-page anonymous manuscript and release manifest; no new license grant.

MANIFEST.json

@@ -1,5 +1,5 @@
 {
-  "release": "flowmo-rebuttal-20260913",
+  "release": "flowmo-icra2027-20260913",
   "original_hub_revision": "069fbace782c99dffcafc7b783f1e2c2acd8ed4e",
   "files": [
     {
@@ -19,13 +19,13 @@
     },
     {
       "path": "METRICS.md",
-      "bytes": 3801,
-      "sha256": "0b05627525b20248053d2c2238840d2b9c41e072eada13adc6e0224d9dcaa252"
+      "bytes": 4981,
+      "sha256": "ba98b13b62de6625fb934344f188cf8264ca45cb4d5aaf6a937faea646b1f653"
     },
     {
       "path": "README.md",
-      "bytes": 8289,
-      "sha256": "86df89aa72429a96b0a99f86a1c4879852ddb398ff22f81470d9faf2e51265f0"
+      "bytes": 8729,
+      "sha256": "589984d8cff811d1a6114a29bd8aceed18c0b7f723eeb884aa237036696d2185"
     },
     {
       "path": "assets/FlowMo_WM_architecture.pptx",
@@ -5134,8 +5134,8 @@
     },
     {
       "path": "paper/FlowMo-WM-ICRA2027.pdf",
-      "bytes": 2087625,
-      "sha256": "8aebaff9bd37371ccc34b0b9472350da6297e561c9dcef671a8ccb9f5795ca3c"
+      "bytes": 2088555,
+      "sha256": "74917dbf1dc0ce62866feb7c569b33170477492cf8f595453200c7f98fec93a4"
     },
     {
       "path": "paper/IEEEtran.bst",
@@ -5189,13 +5189,13 @@
     },
     {
       "path": "paper/main.tex",
-      "bytes": 32471,
-      "sha256": "e473d4c07b0ff1bfd88b08c092ec7167a56db89ae578f3f4eec80716c19857a1"
+      "bytes": 33197,
+      "sha256": "dd98b0e74120cb1fe885f4f73f891d6fbd428239e0ebf08d8192a64250319b43"
     },
     {
       "path": "paper/reference.bib",
-      "bytes": 13371,
-      "sha256": "8576d31183975e0922c3e7e7bf24a542f16d54a08446ccc6adc84b0f022b1c96"
+      "bytes": 13416,
+      "sha256": "86b6aa577b2f1d71a12caa3d724e1ae84a64679ce89f92feedcb9761997b474a"
     },
     {
       "path": "paper/tables/table_planning_summary.tex",
@@ -5738,5 +5738,6 @@
     "archived_original": 3,
     "reconstructed_from_recorded_protocol": 13,
     "exact_metadata_restoration": 2
-  }
+  },
+  "previous_hub_revision": "eb4fc81b9d6e50ade634b6058f8263c3ccb08bae"
 }

METRICS.md

@@ -22,7 +22,11 @@ Paths in the table are relative to the repository or the stated report root. Lar
 
 **Uncertainty.** Table II uses the sample standard deviation across eight independent model seeds. Figure 6A/C/D/E/F use the archived Student-t 95% mean intervals (seven degrees of freedom). Figure 6B uses the archived 50,000-draw seed bootstrap. These error bars are not independent-window confidence intervals.
 
-**Success and conditioning.** Planning success is the task-specific Boolean recorded per episode. Table I final distance uses all episodes; energy and number of steps use successful episodes only. The corrected FlowMo initial count is 3,942/4,000, or 98.55%. Counts come directly from individual records, not rounded rates.
+**Success and conditioning.** Planning success is the task-specific Boolean recorded per episode. The goal radius is 6.5% of workspace width. Reaching requires entry into the goal region; waypoint tasks require completing all goals in order. Station keeping uses terminal occupancy of the goal region and stops early on entry or occupancy from step 101 of a 300-step budget. Continuous residence throughout the episode is not required. Table I final distance uses all episodes; energy and number of steps use successful episodes only. The corrected FlowMo initial count is 3,942/4,000, or 98.55%. Counts come directly from individual records, not rounded rates.
+
+**Planning initialization.** Planning starts from rest, at `(0.5W, 0.5W)` for station keeping and `(0.2W, 0.2W)` for other tasks. Station heading is 0.3 radians; other initial headings are sampled uniformly over `[-pi, pi]`. The common task initialization and parameter/flow sampling rules differ from the fully sampled prediction dataset. Planning actions are generated online by the evaluated controller.
+
+**Velocity probes.** The historical JSON key `momentum` denotes the three-component target `[vx, vy, omega]`: linear and angular velocity. It does not denote `[mass*vx, mass*vy, inertia*omega]`. The historical key and numerical results are retained for compatibility; the manuscript names the actual predicted quantities. The original linear-probe R-squared is 0.667; formal nonlinear means from `c` and `[z,c]` are 0.701 and 0.818.
 
 **Figure 6E condition labels.** R0: reach in no flow; R1: reach in uniform flow; SV: station keeping in a vortex; SD: station keeping in smooth dynamic flow; S2: station keeping with doubled uniform-flow magnitude; ZF: zigzag in a random Fourier field. Each plotted seed summarizes thirty episodes.
 

README.md

@@ -11,6 +11,8 @@ library_name: pytorch
 
 Code, simulation inputs, trained checkpoints, numerical results and figure sources for FlowMo-WM. The [original preprint](https://arxiv.org/abs/2606.13817) and the revised manuscript describe a predictor with a recent-history state, sparse longer-history context and zero-context residual dynamics. This repository contains the original FlowMo study and its completed eight-seed follow-up controls.
 
+The ICRA 2027 submission draft was frozen on 2026-09-13 after clarifying velocity-probe terminology, task-specific success/stopping rules and planning initialization, and adding direct temporal/memory related work. These manuscript and documentation changes preserve all recorded experiments, checkpoints, tables and quantitative figures. The source and reviewed eight-page PDF are in `paper/`; the target definitions are in `METRICS.md`.
+
 The matched H60 comparison supports lower error than Recent-8 and error close to Full-32 with 16 instead of 32 image encodings. Image-encoding count is not a measured end-to-end latency improvement. The LeWorldModel-, RSSM- and TD-MPC2-style comparisons are adaptations to the common supervised pose-prediction interface, not reproductions of complete agents.
 
 ## Rebuild the paper materials

paper/FlowMo-WM-ICRA2027.pdf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8aebaff9bd37371ccc34b0b9472350da6297e561c9dcef671a8ccb9f5795ca3c
-size 2087625
+oid sha256:74917dbf1dc0ce62866feb7c569b33170477492cf8f595453200c7f98fec93a4
+size 2088555

paper/main.tex

@@ -70,7 +70,7 @@ Our contributions are: (i) a visual prediction model that combines recent and sp
 PlaNet, Dreamer, and TD-MPC2 learn latent dynamics for planning or policy improvement \cite{hafner2019planet,hafner2025dreamerv3,hansen2024tdmpc2}. DINO-WM predicts pretrained visual features for goal-directed action optimization \cite{zhou2025dinowm}. Iso-Dream separates controllable and noncontrollable visual dynamics \cite{pan2022isodream}; ActSWM trains action-sensitive rollouts, and DWM regularizes complementary world and action effects \cite{gan2026actswm,zhang2026dwm}. FlowMo-WM focuses on combining recent observations with a sparse longer history in an end-to-end pose predictor.
 
 \paragraph{Temporal inference and unknown dynamics}
-Recurrent latent models recover predictive state under partial observation \cite{karl2017dvbf,han2020vrm}. Online identification, CaDM, and ProtoCAD infer dynamics information from history \cite{yu2017uposi,lee2020cadm,wang2022protocad}. Context-aware Neural ODE control initially uses privileged environmental information and then infers context from state--action history \cite{yu2026contextode}. ContextWM, Denoised MDPs, and ACRO investigate useful representations under changing or exogenous visual information \cite{wu2023contextwm,wang2022denoisedmdp,islam2023acro}. Our setting makes the disturbance invisible while retaining its effect on observed vehicle motion.
+Recurrent latent models recover predictive state under partial observation \cite{karl2017dvbf,han2020vrm}. Online identification, CaDM, and ProtoCAD infer dynamics information from history \cite{yu2017uposi,lee2020cadm,wang2022protocad}. Context-aware Neural ODE control initially uses privileged environmental information and then infers context from state--action history \cite{yu2026contextode}. ContextWM, Denoised MDPs, and ACRO study contextual or exogenous visual information \cite{wu2023contextwm,wang2022denoisedmdp,islam2023acro}. MTS3 couples slow task and fast state dynamics through probabilistic inference \cite{shaj2023mts3}; SF-RSSM coordinates fast and slow branches through distillation \cite{wang2026sfrssm}. WEAVER combines sparse memory and recent history for multiview generation \cite{jain2026weaver}. FlowMo-WM uses a compact history context in each step of a zero-reference residual pose rollout, with matched controls to evaluate this formulation under hidden disturbances.
 
 \paragraph{Motion structure and marine control}
 NewtonianVAE introduces physical structure into visual latent dynamics \cite{jaques2021newtonianvae}. Hydrodynamic models and disturbance estimates also support marine control \cite{fossen2021handbook,fossen2017ilos,belleter2019observer}. FlowMo-WM learns a free latent increment from image--action windows and pose targets. It receives no measured velocity, current, or vessel parameters at inference; the temporal representations are evaluated for their predictive content.
@@ -149,7 +149,7 @@ Both sample mass, inertia, drag, thrust scale, and actuator time constant.
 For batching, actions are stored as a common length-three thruster-command vector in the 2D simulator; the twin boat uses the first two active thruster commands, while the triangular boat uses all three. 
 The two morphologies probe different control structures: the twin boat must turn to redirect thrust, while the triangular boat can compensate for lateral drift more directly.
 Planning tasks include reaching targets, station-keeping, square waypoints, and zigzag waypoints. 
-Reaching tests motion toward a fixed goal; station keeping tests holding under drift; the waypoint tasks require repeated redirection.
+Reaching tests motion toward a fixed goal; station keeping tests regulation around a target under drift; the waypoint tasks require repeated redirection.
 Training and prediction evaluation were run on an NVIDIA RTX 5090 GPU.
 
 \begin{figure*}[!tb]
@@ -162,7 +162,7 @@ Training and prediction evaluation were run on an NVIDIA RTX 5090 GPU.
 \subsection{Training and Evaluation}
 \label{sec:training_evaluation}
 
-Training uses $2{,}400$ episodes of 300 simulator steps, corresponding to $720{,}000$ training transitions and $722{,}400$ simulator states including initial states. The held-out prediction set contains $480$ episodes, or $144{,}000$ transitions. All reported prediction and planning results are computed on held-out episodes with independently sampled initial conditions, physical parameters, action sequences, and flow-field instances; no test rollout windows are drawn from training episodes. 
+Training uses $2{,}400$ episodes of 300 simulator steps, corresponding to $720{,}000$ training transitions and $722{,}400$ simulator states including initial states. The held-out prediction set contains $480$ episodes, or $144{,}000$ transitions. Training and prediction-test episodes independently sample initial conditions, physical parameters, action sequences, and flow-field instances; no test rollout windows are drawn from training episodes. 
 The $160\times160$ RGB observations are rendered online from stored simulator states without overlays. 
 The prediction target is pose $y_t=(x_t,y_t,\cos\theta_t,\sin\theta_t)$; positions are normalized from the $[0,10]\times[0,10]$ workspace during training, while heading is represented by cosine and sine. 
 Episodes vary in vehicle morphology, physical parameters, flow field, and trajectory-generation mode, including random action, push-and-coast, passive drift, goal seeking, and waypoint following. 
@@ -178,12 +178,13 @@ Parameter counts are closely matched: FlowMo-WM has $663{,}964$ parameters, LeWo
 Prediction evaluation uses $16{,}384$ open-loop 60-step rollout windows with ground-truth future actions. 
 Spatial error is reported as $E_{\rm pos}(h)=100\|\hat p_{t+h}-p_{t+h}\|_2/W$, averaged over evaluated windows after decoding positions to workspace coordinates. Heading error from predicted and target sine--cosine headings is a secondary diagnostic; the main figures emphasize position error and task success. 
 For prediction-time context diagnostics, we also evaluate FlowMo-WM with normal context, zero context $c_t=\mathbf{0}$, and shuffled context from another sample. 
-Frozen linear probes~\cite{alain2016understanding} are fit after training from $z_t$, $c_t$, and $[z_t,c_t]$ to privileged momentum and flow quantities only for analysis.
+Frozen linear probes~\cite{alain2016understanding} are fit after training from $z_t$, $c_t$, and $[z_t,c_t]$ to linear and angular velocity $(v_x,v_y,\omega)$ and flow quantities only for analysis.
 
 For downstream planning, learned world models score sampled candidate action sequences through latent rollout with the same route-aware planner. 
 Each planning update evaluates 512 sequences over a 45-step horizon for four refinement rounds and updates the sampling distribution from the 64 best-scoring sequences.
 We run four tasks, two boats, ten flow families, and 50 episodes per setting. 
-Success means reaching the active goal within a radius equal to $6.5\%$ of workspace width; final distance $d$ is measured to the active terminal goal and reported as $100d/W$; the reported energy is the action-effort proxy $\sum_t\|a_t\|_2^2$ over successful episodes. 
+Planning starts from rest at $(0.5W,0.5W)$ for station keeping and $(0.2W,0.2W)$ otherwise; heading is fixed at $0.3$ rad for station keeping and sampled uniformly otherwise. All methods share task initializations and parameter/flow sampling rules, with commands generated online.
+The goal radius is $6.5\%$ of workspace width. Reaching succeeds upon entry; waypoints require ordered completion. Station-keeping success uses terminal occupancy of this radius, with early stopping on entry or occupancy from step 101 of a 300-step budget. Final distance $d$ is to the active terminal goal, reported as $100d/W$; energy is the action-effort proxy $\sum_t\|a_t\|_2^2$ over successful episodes. 
 We also include LOS references: PID/LOS~\cite{fossen2021handbook}, No-Flow LOS, Current-Estimator LOS, and Oracle-Flow LOS. 
 These are downstream references with different information assumptions, not direct learned visual world-model comparison methods.
 PID/LOS uses hand-designed line-of-sight tracking from pose and goal, No-Flow LOS ignores ambient transport, Current-Estimator LOS uses a recent-displacement heuristic, and Oracle-Flow LOS uses privileged simulator flow as a diagnostic feed-forward signal.
@@ -235,9 +236,9 @@ The shuffled-context condition tests sensitivity to context mismatch. These infe
 
 \paragraph{Representation analysis}
 The context ablations test whether the long-history pathway is functionally used during rollout; frozen probes provide a complementary view of what information is present in the learned factors. 
-Momentum is linearly decodable from $[z_t,c_t]$ with $R^2=0.667$, indicating that the learned representation contains substantial motion-state information even though velocity is never used as a training target. 
+Linear and angular velocity are jointly decodable from $[z_t,c_t]$ with linear-probe $R^2=0.667$, indicating substantial motion-state information without direct velocity supervision. 
 Local flow and episode-level drift have near-zero linear $R^2$, which suggests that the context improves prediction through a task-useful latent code rather than by storing a trivially linear physical current vector. 
-The formal nonlinear probes recover momentum from $c$ and $[z,c]$ with $R^2=0.701$ and $0.818$, respectively, and predict boat type from $c$ with balanced accuracy $0.904$. Expanded flow probes provide no clear evidence for an identifiable local, episode-average, or history-averaged flow representation. The context therefore mixes predictive dynamics information; context use does not establish physical disentanglement.
+The formal nonlinear probes recover these velocities from $c$ and $[z,c]$ with $R^2=0.701$ and $0.818$, respectively, and predict boat type from $c$ with balanced accuracy $0.904$. Expanded flow probes provide no clear evidence for an identifiable local, episode-average, or history-averaged flow representation. The context therefore mixes predictive dynamics information; context use does not establish physical disentanglement.
 
 \paragraph{Planning results}
 Table~\ref{tab:planning_summary} reports aggregate planning results split by boat morphology. 

paper/reference.bib

@@ -11,17 +11,17 @@
 @inproceedings{hafner2019planet,
   title     = {Learning Latent Dynamics for Planning from Pixels},
   author    = {Hafner, Danijar and Lillicrap, Timothy and Fischer, Ian and Villegas, Ruben and Ha, David and Lee, Honglak and Davidson, James},
-  booktitle = {Proceedings of the 36th International Conference on Machine Learning},
+  booktitle = {Proc. Int. Conf. Mach. Learn. (ICML)},
   pages     = {2555--2565},
   year      = {2019},
   volume    = {97},
-  series    = {Proceedings of Machine Learning Research}
+  series    = {PMLR}
 }
 
 @inproceedings{hafner2020dreamer,
   title     = {Dream to Control: Learning Behaviors by Latent Imagination},
   author    = {Hafner, Danijar and Lillicrap, Timothy and Ba, Jimmy and Norouzi, Mohammad},
-  booktitle = {International Conference on Learning Representations},
+  booktitle = {Int. Conf. Learn. Represent. (ICLR)},
   year      = {2020}
 }
 
@@ -35,18 +35,18 @@
 @inproceedings{hansen2024tdmpc2,
   title     = {{TD-MPC2}: Scalable, Robust World Models for Continuous Control},
   author    = {Hansen, Nicklas and Su, Hao and Wang, Xiaolong},
-  booktitle = {International Conference on Learning Representations},
+  booktitle = {Int. Conf. Learn. Represent. (ICLR)},
   year      = {2024}
 }
 
 @inproceedings{wu2023daydreamer,
   title     = {{DayDreamer}: World Models for Physical Robot Learning},
   author    = {Wu, Philipp and Escontrela, Alejandro and Hafner, Danijar and Abbeel, Pieter and Goldberg, Ken},
-  booktitle = {Proceedings of The 6th Conference on Robot Learning},
+  booktitle = {Proc. Conf. Robot Learn. (CoRL)},
   pages     = {2226--2240},
   year      = {2023},
   volume    = {205},
-  series    = {Proceedings of Machine Learning Research}
+  series    = {PMLR}
 }
 
 @article{fossen2017ilos,
@@ -84,14 +84,14 @@
 @inproceedings{karl2017dvbf,
   title     = {Deep Variational Bayes Filters: Unsupervised Learning of State Space Models from Raw Data},
   author    = {Karl, Maximilian and Soelch, Maximilian and Bayer, Justin and van der Smagt, Patrick},
-  booktitle = {International Conference on Learning Representations},
+  booktitle = {Int. Conf. Learn. Represent. (ICLR)},
   year      = {2017}
 }
 
 @inproceedings{han2020vrm,
   title     = {Variational Recurrent Models for Solving Partially Observable Control Tasks},
   author    = {Han, Dongqi and Doya, Kenji and Tani, Jun},
-  booktitle = {International Conference on Learning Representations},
+  booktitle = {Int. Conf. Learn. Represent. (ICLR)},
   year      = {2020}
 }
 
@@ -134,24 +134,24 @@
 @inproceedings{rakelly2019pearl,
   title     = {Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Variables},
   author    = {Rakelly, Kate and Zhou, Aurick and Finn, Chelsea and Levine, Sergey and Quillen, Deirdre},
-  booktitle = {Proceedings of the 36th International Conference on Machine Learning},
+  booktitle = {Proc. Int. Conf. Mach. Learn. (ICML)},
   pages     = {5331--5340},
   year      = {2019},
   volume    = {97},
-  series    = {Proceedings of Machine Learning Research}
+  series    = {PMLR}
 }
 
 @inproceedings{zintgraf2020varibad,
   title     = {{VariBAD}: A Very Good Method for Bayes-Adaptive Deep {RL} via Meta-Learning},
   author    = {Zintgraf, Luisa and Shiarlis, Kyriacos and Igl, Maximilian and Schulze, Sebastian and Gal, Yarin and Hofmann, Katja and Whiteson, Shimon},
-  booktitle = {International Conference on Learning Representations},
+  booktitle = {Int. Conf. Learn. Represent. (ICLR)},
   year      = {2020}
 }
 
 @inproceedings{wu2023contextwm,
   title     = {Pre-training Contextualized World Models with In-the-wild Videos for Reinforcement Learning},
   author    = {Wu, Jialong and Ma, Haoyu and Deng, Chaoyi and Long, Mingsheng},
-  booktitle = {Advances in Neural Information Processing Systems},
+  booktitle = {Adv. Neural Inf. Process. Syst.},
   volume    = {36},
   year      = {2023}
 }
@@ -159,79 +159,76 @@
 @inproceedings{wang2022denoisedmdp,
   title     = {Denoised {MDPs}: Learning World Models Better Than the World Itself},
   author    = {Wang, Tongzhou and Du, Simon and Torralba, Antonio and Isola, Phillip and Zhang, Amy and Tian, Yuandong},
-  booktitle = {Proceedings of the 39th International Conference on Machine Learning},
+  booktitle = {Proc. Int. Conf. Mach. Learn. (ICML)},
   pages     = {22591--22612},
   year      = {2022},
   volume    = {162},
-  series    = {Proceedings of Machine Learning Research}
+  series    = {PMLR}
 }
 
 @inproceedings{islam2023acro,
   title     = {Principled Offline {RL} in the Presence of Rich Exogenous Information},
   author    = {Islam, Riashat and Tomar, Manan and Lamb, Alex and Efroni, Yonathan and Zang, Hongyu and Didolkar, Aniket Rajiv and Misra, Dipendra and Li, Xin and Van Seijen, Harm and Tachet Des Combes, Remi and Langford, John},
-  booktitle = {Proceedings of the 40th International Conference on Machine Learning},
+  booktitle = {Proc. Int. Conf. Mach. Learn. (ICML)},
   pages     = {14390--14421},
   year      = {2023},
   volume    = {202},
-  series    = {Proceedings of Machine Learning Research}
+  series    = {PMLR}
 }
 
 @inproceedings{zhou2025dinowm,
  title={{DINO-WM}: World Models on Pre-trained Visual Features enable Zero-shot Planning},
  author={Zhou, Gaoyue and Pan, Hengkai and LeCun, Yann and Pinto, Lerrel},
- booktitle={International Conference on Machine Learning}, volume={267},
- series={Proceedings of Machine Learning Research}, pages={79115--79135}, year={2025},
- url={https://proceedings.mlr.press/v267/zhou25t.html}
+ booktitle={Proc. Int. Conf. Mach. Learn. (ICML)}, volume={267},
+ series={PMLR}, pages={79115--79135}, year={2025}
 }
 
 @inproceedings{pan2022isodream,
  title={{Iso-Dream}: Isolating and Leveraging Noncontrollable Visual Dynamics in World Models},
  author={Pan, Minting and Zhu, Xiangming and Wang, Yunbo and Yang, Xiaokang},
- booktitle={Advances in Neural Information Processing Systems}, volume={35}, year={2022},
- url={https://proceedings.neurips.cc/paper_files/paper/2022/hash/9316769afaaeeaad42a9e3633b14e801-Abstract-Conference.html}
+ booktitle={Adv. Neural Inf. Process. Syst.}, volume={35}, year={2022}
 }
 
 @article{gan2026actswm,
  title={{ActSWM}: Action-Sensitive World Models for Long-Horizon Planning in Open-World Games},
  author={Gan, Zhenfeng and Zeng, ZiTong and Cheng, Jiajun and Song, Yeke and Tang, Yongyi and Wang, Xueqian},
  journal={arXiv preprint arXiv:2607.26712}, year={2026},
- doi={10.48550/arXiv.2607.26712}, url={https://arxiv.org/abs/2607.26712}
+ doi={10.48550/arXiv.2607.26712}
 }
 
 @article{zhang2026dwm,
  title={{DWM}: Separating World Effects from Actions in Latent World Models},
  author={Zhang, Yi-Ge and Du, Tianqi and Zhang, Qi and Wang, Yisen},
  journal={arXiv preprint arXiv:2607.18715}, year={2026},
- doi={10.48550/arXiv.2607.18715}, url={https://arxiv.org/abs/2607.18715}
+ doi={10.48550/arXiv.2607.18715}
 }
 
 @inproceedings{lee2020cadm,
   title={Context-aware Dynamics Model for Generalization in Model-Based Reinforcement Learning},
   author={Lee, Kimin and Seo, Younggyo and Lee, Seunghyun and Lee, Honglak and Shin, Jinwoo},
-  booktitle={International Conference on Machine Learning},
-  pages={5757--5766}, volume={119}, publisher={PMLR}, year={2020},
-  url={https://proceedings.mlr.press/v119/lee20g.html}
+  booktitle={Proc. Int. Conf. Mach. Learn. (ICML)},
+  pages={5757--5766}, volume={119}, publisher={PMLR}, year={2020}
 }
 
 @article{wang2022protocad,
  title={Prototypical Context-Aware Dynamics Generalization for High-Dimensional Model-Based Reinforcement Learning},
  author={Wang, Junjie and Mu, Yao and Li, Dong and Zhang, Qichao and Zhao, Dongbin and Zhuang, Yuzheng and Luo, Ping and Wang, Bin and Hao, Jianye},
  journal={arXiv preprint arXiv:2211.12774}, year={2022},
- doi={10.48550/arXiv.2211.12774}, url={https://arxiv.org/abs/2211.12774}
+ doi={10.48550/arXiv.2211.12774}
 }
 
 @article{yu2026contextode,
  title={Learning Context-Aware Neural ODE Dynamics for Adaptive Robotic Control},
  author={Yu, Shao-Yi and Wang, Jen-Wei and Horii, Maya and Tomizuka, Masayoshi and Garg, Vikas},
  journal={arXiv preprint arXiv:2606.15469}, year={2026},
- doi={10.48550/arXiv.2606.15469}, url={https://arxiv.org/abs/2606.15469}
+ doi={10.48550/arXiv.2606.15469}
 }
 
 @inproceedings{jaques2021newtonianvae,
  title={{NewtonianVAE}: Proportional Control and Goal Identification from Pixels via Physical Latent Spaces},
  author={Jaques, Miguel and Burke, Michael and Hospedales, Timothy},
  booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition},
- year={2021}, url={https://arxiv.org/abs/2006.01959}
+ year={2021}
 }
 
 @book{fossen2021handbook,
@@ -246,7 +243,7 @@
 @inproceedings{liu2018coordconv,
   title     = {An Intriguing Failing of Convolutional Neural Networks and the {CoordConv} Solution},
   author    = {Liu, Rosanne and Lehman, Joel and Molino, Piero and Such, Felipe Petroski and Frank, Eric and Sergeev, Alex and Yosinski, Jason},
-  booktitle = {Advances in Neural Information Processing Systems},
+  booktitle = {Adv. Neural Inf. Process. Syst.},
   volume    = {31},
   year      = {2018}
 }
@@ -271,18 +268,18 @@
 @inproceedings{loshchilov2019decoupled,
   title     = {Decoupled Weight Decay Regularization},
   author    = {Loshchilov, Ilya and Hutter, Frank},
-  booktitle = {International Conference on Learning Representations},
+  booktitle = {Int. Conf. Learn. Represent. (ICLR)},
   year      = {2019}
 }
 
 @inproceedings{hansen2022tdmpc,
   title     = {Temporal Difference Learning for Model Predictive Control},
   author    = {Hansen, Nicklas A. and Su, Hao and Wang, Xiaolong},
-  booktitle = {Proceedings of the 39th International Conference on Machine Learning},
+  booktitle = {Proc. Int. Conf. Mach. Learn. (ICML)},
   pages     = {8387--8406},
   year      = {2022},
   volume    = {162},
-  series    = {Proceedings of Machine Learning Research}
+  series    = {PMLR}
 }
 
 @article{alain2016understanding,
@@ -302,3 +299,31 @@
   year    = {2019},
   doi     = {10.1016/j.automatica.2018.11.008}
 }
+
+@inproceedings{shaj2023mts3,
+  title     = {Multi Time Scale World Models},
+  author    = {Shaj, Vaisakh and Gholam Zadeh, Saleh and Demir, Ozan and Douat, Luiz Ricardo and Neumann, Gerhard},
+  booktitle = {Adv. Neural Inf. Process. Syst.},
+  volume    = {36},
+  year      = {2023},
+  doi       = {10.52202/075280-1163}
+}
+
+@inproceedings{wang2026sfrssm,
+  title     = {Beyond Single-Speed Reasoning: Coordinating Fast and Slow Dynamics for Efficient World Modeling},
+  author    = {Wang, Hongwei and Huang, Yangru and Chen, Guangyao and Wang, Xu and Jin, Yi},
+  booktitle = {Proc. AAAI Conf. Artif. Intell.},
+  volume    = {40},
+  number    = {31},
+  pages     = {26215--26223},
+  year      = {2026},
+  doi       = {10.1609/aaai.v40i31.39825}
+}
+
+@article{jain2026weaver,
+  title   = {{WEAVER}, Better, Faster, Longer: An Effective World Model for Robotic Manipulation},
+  author  = {Jain, Arnav Kumar and Wu, Yilin and Farebrother, Jesse and Swamy, Gokul and Bajcsy, Andrea},
+  journal = {arXiv preprint arXiv:2606.13672},
+  year    = {2026},
+  doi     = {10.48550/arXiv.2606.13672}
+}