acceptodds
Withdrawn from ICLR 2027

From Contraction to Chaos: Compounding Error Resistant World Models for Robots

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Learned world models are usually rolled out autoregressively: each prediction is fed back as input, so small one-step errors compound over the horizon. How fast they compound is determined by the system’s sensitivity to initial conditions, one of the three mathematical properties of chaos. Under a tracking controller, deviations are corrected, and errors stay bounded; in a chaotic system, they grow exponentially. Auto-regressive (AR) models, such as the temporal convolutional networks widely used in mobile robotics, learn a simple one-step map but pay the full cost of compounding. One-shot Multi-step (MS) models never feed back their own errors, which theory predicts should pay off over long horizons. However, they cannot reuse what they learn for one step when predicting the next, so they need more data to learn the multi-step map. Neither class dominates across the sensitivity spectrum. We propose Multi-Step Temporal Mixture Probabilistic (MTM-Pro), a hybrid world model trained offline on logged trajectories with two coupled objectives: an MS objective that fuses overlapping multi-step forecasts through a learned temporal mixture, and an AR objective that rolls the model out on its own predictions. We evaluate on five datasets chosen to span the sensitivity spectrum: closed-loop quadcopter tracking on nominal trajectories, on aggressive trajectories, and at the edge of the control envelope; open-loop driving of an off-road uncrewed ground vehicle (UGV); and the chaotic Lorenz system. AR baselines degrade as sensitivity increases, while MS baselines become competitive only on the most sensitive system. MTM-Pro matches the best baseline on the least sensitive datasets and reduces 5-second drift by more than three orders of magnitude on Lorenz, making it the only model reliable across the whole spectrum. Our code and links to all datasets are publicly available at https://anonymous.4open.science/r/MTM-Pro.

Primary area:
applications to robotics, autonomy, planning
Keywords:
robot learning, world models, learned dynamics models, multi-step prediction, compounding error, probabilistic forecasting, mixture models, long-horizon prediction, distribution shift

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_XnmZOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper focuses on compounding error in learned dynamics models for robots. The authors relate how fast the error grows to how sensitive the system is to initial conditions, and propose MTM-Pro, which is trained with a multi-step loss (overlapping forecasts fused by a learned mixture) and a rollout loss, but runs as a one-step AR model at test time. Results show 5s of open-loop drift on four robot datasets and the Lorenz system.

Critical strengths
  1. The sensitivity view is a clear framing that explains why AR and multi-step models win in different regimes.

  2. Multi-step supervision is used only in training, so deployment is a single forward pass (about 180 Hz on an embedded CPU).

  3. Baselines are compared at matched parameter budgets; the appendix is thorough, and code and a new off-road UGV dataset are released.

Critical weaknesses
  1. I am not fully convinced by the motivation. As far as I know, a world model on a real robot gets ground-truth observations or state estimates all the time, and these correct the drift. MPC also replans at every step with a horizon well below 5 s. When does a robot actually need to predict 5 s ahead with no feedback at all? Some concrete cases would help.

  2. The presentation of this paper needs to be improved. The introduction is under-cited. 1) Terms like world model, MPC, and MBDRL appear without any reference. 2) Many paragraphs are very long, which makes the paper hard to follow. 3) Some numbers do not match across the paper. AR TCN training time on PI-TCN is 23 h in Table 1 but 53 h in Sec. 4.4. Lorenz drift reduction is "more than three orders of magnitude" in the abstract but "more than four" in Sec. 4.1. I found a few more. Please go through all the reported numbers carefully.

  3. There is no closed-loop real-world experiment. Everything is open-loop drift on logged data, although the paper is motivated by MPC and MBRL. The authors mention this in the limitations, but I still think at least one MPC experiment is needed, even in simulation, to show that lower drift actually helps control.

Questions for authors

See Weaknesses

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

This paper proposes MTM-Pro, a world model that combines autoregressive and multi-step prediction objectives to reduce long-horizon error accumulation. Overrall, I think this paper has huge shortcomings in terms of exps and novelty.

Critical strengths

None

Critical weaknesses
  1. The Lorenz test set includes a trajectory with the same dynamics parameters and initial conditions as training, making the generalization results questionable. The PI-TCN difficulty split is also based on baseline prediction errors rather than independently measured dynamical properties.

  2. This paper lacks convincing ablations separating the effects of the temporal mixture and the two training objectives. Moreover, the claimed sensitivity spectrum is not quantitatively established, making it difficult to attribute the observed performance differences to the proposed explanation.

Questions for authors
  1. Why does the Lorenz evaluation reuse training dynamics parameters and initial conditions?

  2. How was the PI-TCN difficulty split validated independently of baseline prediction errors?

  3. Can the authors provide separate ablations for the temporal mixture and each training objective?

Reviewer_AX43OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies compounding error in learned robot dynamics models over long open-loop rollouts. The authors claim that Autoregressive (AR) models should work best on contracting systems, while one-shot multi-step (MS) models should work best on chaotic systems. The authors propose MTM-Pro, an hybrid probabilistic model trained with two objectives: an MS loss and a AR compounded-prediction loss. In experiments, MTM-Pro is reported as best on all datasets compared to other AR and MS baselines.

Critical strengths
  1. The authors study an important practical problem in model-based control and planning for robotics. Compounding multi-step prediction error in learned dynamics models is a key bottleneck in robotics, and analyzing model behavior across different dynamical regimes provides an intuitive and motivating perspective.
  2. The experimental results are promising. MTM-Pro is competitive across every dataset and achieves large improvements in regimes where error compounding is severe, such as on the Lorenz dataset.
Critical weaknesses
  1. Lack of Ablation Studies on Training vs. Deployment Mechanisms. The absence of ablation studies makes it unclear how effective the individual mechanisms are. For instance, MTM-Pro uses both MS and AR (CP) losses during training, yet at deployment it relies only on 1-step AR predictions for rollouts. Why not use the MS predictions or fuse both predictions at deployment to achieve better results?
  2. There is insufficient justification for the various loss terms and weighting schemes. In Section 3 ("Objective, training and deployment"), the objective also incorporates auto-weighting, orientation residuals, and log-variance bound regularization. Ablation studies are needed to justify the contribution of each component.
  3. Limited Novelty and Missing Related Work: The main components are close to existing ideas. Combining recursive and direct multi-step forecasting is well studied in time-series forecasting, such as the DirRec strategy [1] and the review of hybrid recursive/direct strategies in [2]. The claim of being "the first model to learn a mixture of the two" should be contextualized and discussed against these prior works.
  4. Minor (Presentation Clarity): The temporal mixture formulation in Section 3.1 is dense and hard to follow. I recommend adding a flowchart or system framework diagram to illustrate the dataflow.

[1] Sorjamaa and Lendasse, "Time Series Prediction Using DirRec Strategy," ESANN, 2006.
[2] Ben Taieb et al., "A Review and Comparison of Strategies for Multi-Step Ahead Time Series Forecasting Based on the NN5 Forecasting Competition," Expert Systems with Applications, 2012.

Questions for authors
  1. Since MTM-Pro is trained with both MS and CP losses, why does deployment only use the 1-step AR prediction for rollouts rather than using the MS predictions or fusing both? Could you provide an ablation study comparing these deployment choices?
  2. Could you provide ablation studies to isolate the individual contributions of the MS loss, CP loss, orientation loss, auto-weighting, and log-variance bound regularization?
  3. How does MTM-Pro conceptually differ from classic hybrid direct/recursive multi-step forecasting strategies in the time-series literature (e.g., DirRec [1] and [2])? Please discuss these works and qualify the novelty claim in Section 2.

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

Related papers

Discussion (0)

Sign in to comment.