acceptodds
Under review as a conference paper at ICLR 2027

EcoLatent: A Physics Guided Latent Inner World Model for Vehicles and Robots

Abstract

World models provide a basis for embodied decision-making, yet modeling thed surrounding environment alone leaves an essential question unanswered: how will the agent's own internal world evolve under its actions? We introduce EcoLATENT, a physics-guided latent inner world model for vehicles and robots, and instantiate it on vehicle longitudinal and powertrain dynamics. The model combines explicit physical state with a recurrent latent operating context inferred from causal sensor history. Physical equations and learned residual dynamics jointly advance the internal state under candidate actions, enabling imagined trajectories to support downstream decisions. We use vehicle energy management as a concrete case study, with energy represented as one component of the internal world. On 842 h of operational vehicle logs, an 85k-parameter model achieves 0.653 m/s speed RMSE and 1.61% cumulative-fuel normalized MAE over 60 s, matching a 272k-parameter GRU on speed while preserving a structural fuel-cut rule. A policy is optimized through the learned inner world; cross-model evaluation reveals that its estimated benefit depends strongly on the evaluation dynamics. These results demonstrate a compact inner-world representation and an imagination-based decision pipeline in vehicles, while motivating extensions to other internal subsystems and robotic embodiments.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Latent World Model, Vehicle CAN Data, Physics-Guided Learning, Counterfactual Rollout, Energy-Efficient Control, Vehicle Energy Management

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_9aNZOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies how to predict motion, powertrain states, and fuel consumption from a vehicle’s recent operating history and future actions, and uses these predictions to optimize energy-saving strategies. The method encodes historical observations with a GRU, combines latent operating states with explicit physical states, recursively predicts the next 60 seconds using physical equations and neural-network residuals, and finally trains a policy through a frozen differentiable model.

On two vehicles and 842 hours of logs, EcoLatent with 85k parameters reaches speed prediction accuracy close to that of a 272k-parameter GRU: RMSEs of 0.653 and 0.656 m/s, respectively, and cumulative fuel nMAEs of 1.61% and 1.65%, respectively (Table 1). The results support that the model is more compact than the selected baseline and slightly improves fuel prediction, but it shows no significant advantage in speed accuracy.

It is commendable that the paper conducts cross-model policy evaluation and explicitly reports benefit decay and degradation under difficult operating conditions. These results help assess the practical limitations of the prediction model. However, the machine learning contributions still have the following problems:

  1. Insufficient methodological novelty. The core consists of recurrent encoding of history, physics residual learning, and differentiable policy optimization, all of which have established precedents. The paper has not yet demonstrated that this combination produces a unique learning mechanism or an important new capability. The term “internal world model” also does not correspond to a clearly defined new learning problem.

  2. Insufficient evidence for structural advantages. A capacity-matched GRU baseline is missing; the model without latent context has only 18k parameters, so the ablation confounds capacity changes. Removing physics only increases speed RMSE from 0.653 to 0.674, whereas removing road preview increases it to 1.20 (p. 6). Therefore, the current results cannot yet be fully attributed to the proposed physics and latent-state structure.

  3. Decision effectiveness has not been established. The policy’s estimated energy savings fall from 5.94% in the optimization model to 0.57% in the physical model, and in the latter it is inferior to physical MPC (Table 2). This does not prove that the physical model is ground truth, but it shows that the benefits depend on the evaluation dynamics. The paper exposes the gap between prediction and decision-making, but it has not yet produced a new error explanation, diagnostic method, or solution.

  4. Generalizable empirical conclusions are limited. The tests mainly cover roads repeatedly traversed during training, and the cross-vehicle experiment involves only one held-out vehicle; hilly and long-uphill conditions also do not consistently outperform the GRU. Therefore, they are insufficient to support general generalization or data-efficiency advantages.

The key to the acceptance decision is new machine learning knowledge and the strength of its evidence. The current method offers limited increment, the attribution of its prediction advantages is unclear, and its decision-making contribution has not been independently validated, so rejection is recommended. Capacity-matched experiments and independent action-response validation are most likely to change this judgment.

Critical strengths

It is commendable that the paper conducts cross-model policy evaluation and explicitly reports benefit decay and degradation under difficult operating conditions. These results help assess the practical limitations of the prediction model. However, the machine learning contributions still have the following problems:

Critical weaknesses
  1. Insufficient methodological novelty. The core consists of recurrent encoding of history, physics residual learning, and differentiable policy optimization, all of which have established precedents. The paper has not yet demonstrated that this combination produces a unique learning mechanism or an important new capability. The term “internal world model” also does not correspond to a clearly defined new learning problem.

  2. Insufficient evidence for structural advantages. A capacity-matched GRU baseline is missing; the model without latent context has only 18k parameters, so the ablation confounds capacity changes. Removing physics only increases speed RMSE from 0.653 to 0.674, whereas removing road preview increases it to 1.20 (p. 6). Therefore, the current results cannot yet be fully attributed to the proposed physics and latent-state structure.

  3. Decision effectiveness has not been established. The policy’s estimated energy savings fall from 5.94% in the optimization model to 0.57% in the physical model, and in the latter it is inferior to physical MPC (Table 2). This does not prove that the physical model is ground truth, but it shows that the benefits depend on the evaluation dynamics. The paper exposes the gap between prediction and decision-making, but it has not yet produced a new error explanation, diagnostic method, or solution.

  4. Generalizable empirical conclusions are limited. The tests mainly cover roads repeatedly traversed during training, and the cross-vehicle experiment involves only one held-out vehicle; hilly and long-uphill conditions also do not consistently outperform the GRU. Therefore, they are insufficient to support general generalization or data-efficiency advantages.

Questions for authors
  1. Insufficient methodological novelty. The core consists of recurrent encoding of history, physics residual learning, and differentiable policy optimization, all of which have established precedents. The paper has not yet demonstrated that this combination produces a unique learning mechanism or an important new capability. The term “internal world model” also does not correspond to a clearly defined new learning problem.

  2. Insufficient evidence for structural advantages. A capacity-matched GRU baseline is missing; the model without latent context has only 18k parameters, so the ablation confounds capacity changes. Removing physics only increases speed RMSE from 0.653 to 0.674, whereas removing road preview increases it to 1.20 (p. 6). Therefore, the current results cannot yet be fully attributed to the proposed physics and latent-state structure.

  3. Decision effectiveness has not been established. The policy’s estimated energy savings fall from 5.94% in the optimization model to 0.57% in the physical model, and in the latter it is inferior to physical MPC (Table 2). This does not prove that the physical model is ground truth, but it shows that the benefits depend on the evaluation dynamics. The paper exposes the gap between prediction and decision-making, but it has not yet produced a new error explanation, diagnostic method, or solution.

  4. Generalizable empirical conclusions are limited. The tests mainly cover roads repeatedly traversed during training, and the cross-vehicle experiment involves only one held-out vehicle; hilly and long-uphill conditions also do not consistently outperform the GRU. Therefore, they are insufficient to support general generalization or data-efficiency advantages.

Reviewer_WYN3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents EcoLatent, a physics-guided latent inner world model for predicting the internal dynamics of embodied systems. The model combines physical equations, recurrent latent states, and learned residual dynamics to predict vehicle motion and energy consumption. It also uses imagined rollouts to optimize a downstream energy-management policy. Experiments on 842 hours of real vehicle data show that EcoLatent achieves comparable prediction accuracy to a larger GRU model with fewer parameters. However, the estimated energy savings vary considerably across different dynamics models, highlighting the challenges of using learned dynamics for decision-making.

Critical strengths

The paper addresses an important practical problem by focusing on vehicle energy efficiency and energy management. It combines physical dynamics with learned latent representations in a compact model. The experiments use extensive real-world vehicle data and include some baselines, ablations, and cross-model evaluations. I also appreciate the discussion of negative results, particularly the differences in estimated energy savings across evaluation models.

Critical weaknesses
  1. Similar ideas combining physical models with neural networks and residual learning have been explored in previous studies, including Physical enhanced residual learning (PERL) framework for vehicle trajectory prediction (Long et al., 2025), Hybrid modeling for vehicle lateral dynamics via AGRU with a dual-attention mechanism under limited data (Chen et al., 2024), and KNODE-MPC: A Knowledge-Based Data-Driven Predictive Control Framework for Aerial Robots (Chee et al., 2022). These works are not discussed, and the authors should clarify what is technically new in EcoLatent beyond the proposed inner-world formulation.
  2. Missing comparisons with stronger baselines. The experiments mainly compare EcoLatent with standard physics-based models and neural networks. Comparisons with existing physics-residual learning methods would be more convincing.
  3. The evaluation uses only two vehicles, and most test samples come from roads already covered by the training data. Since road preview has a large impact on performance, testing on unseen routes would be important for demonstrating generalization.
  4. I suggest revising "for Vehicles and Robots" in the title, as all experiments focus on vehicle longitudinal dynamics and energy management. It remains unclear how well the proposed framework would generalize to other internal systems or robotic platforms.
Questions for authors

None

Reviewer_JauYOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents EcoLatent, a model for predicting a vehicle’s internal operating state and evaluating different control actions. It combines physical equations for longitudinal motion and the powertrain with a recurrent network. The network extracts operating context from current and past sensor readings. Physical equations and learned residuals then predict motion, engine state, and cumulative fuel use. The experiments use 842 hours of logs from two vehicles. They test prediction under recorded actions, transfer between vehicles, physical consistency, and policy optimization. The paper also evaluates the same policy with different dynamics models and finds that estimated energy savings vary substantially across models.

Critical strengths

The authors are careful about what the experiments show. Trips are split chronologically, all predictors use the same action and road inputs, and the experiments use three training seeds. Confidence intervals are estimated by bootstrapping trips, accounting for overlapping evaluation windows. EcoLatent has only 85,000 parameters. Its speed prediction is close to that of a 272,000-parameter GRU, and its cumulative fuel prediction is slightly better. The authors also report cases where transfer works poorly, accounting problems in earlier policy evaluations, and large differences between evaluation models. These results help explain where the model can and cannot be used.

Critical weaknesses

The current results mainly show that the model predicts reasonably well on routes driven repeatedly. Its ability to evaluate new control actions still needs validation.

  1. The estimated improvement is 5.94% in the model used to optimize the policy, but only 0.57% in the calibrated physics model. On a separate set of 60 selected free-flow scenarios, the policy uses 2.61% more equivalent energy than replaying the recorded driving actions, while physics-based MPC uses 2.27% less. These differences do not show which evaluator is correct, but they leave the energy benefit unresolved. The paper also does not establish whether the logged ECU torque can be treated as an independently controllable input. The authors acknowledge these limits. To use the model for control, responses to changed actions should be checked through controlled actuator perturbations or an independently validated dynamics model.
  2. Combining physical equations, learned residuals, and a recurrent network is close to the existing hybrid system-identification methods discussed in the paper. The 272,000-parameter GRU has almost the same speed RMSE as EcoLatent. Removing the physics component gives an RMSE of 0.674 m/s, compared with 0.653 m/s for EcoLatent, which is a small difference. The PhysRes comparison changes both how history is used and the model size. A recurrent baseline of similar size, or an accuracy-versus-size comparison, would help show what the proposed structure adds. The small improvement in fuel prediction is supported, but it does not by itself support broader claims about an internal world model.
  3. Removing road preview raises speed RMSE from 0.653 to 1.20 m/s. In addition, 94% of moving test samples fall in map cells visited by at least three training trips, and evaluation windows require at least 90% map coverage of the road ahead. The fitted effective-grade map may also absorb resistance that is specific to a location but is not explicitly modeled. Transfer between vehicles is useful, but it does not establish performance on new routes or routes with incomplete maps. Training and testing on separate routes, together with tests at lower map coverage, would clarify this limitation.
  4. Fuel labels come from an ECU proxy, mass and gear use estimates from the data provider, and the dynamics model uses fitted effective parameters. The model enforces zero predicted fuel use when reported torque is zero. Yet on 7.8% of these steps, the power represented by the fuel proxy still exceeds 5 kW. This shows that the model follows the chosen rule, but does not establish that the rule matches actual fuel use. The paper acknowledges this difference. Independent fuel measurements, or tests with alternative fuel-cut rules and calibration assumptions, would help show whether the prediction and policy results are reliable.
Questions for authors

Q1. Can responses to changed control actions be checked through controlled actuator perturbations or an independently validated simulator? How should the energy benefit be judged when different evaluation models rank the policies differently? Q2. With the same inputs and losses, how does EcoLatent compare with a recurrent predictor of similar size and a recurrent hybrid model? Which component provides the main improvement? Q3. How does the model perform on routes excluded from map construction, or with less map coverage ahead? Can testing on separate routes distinguish transferable dynamics from compensation for specific locations? Q4. Can independent measurements validate the fuel proxy and fuel-cut rule? How much do prediction results and policy rankings change under different fuel-cut rules or calibration assumptions?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.