Pith. sign in

REVIEW 2 major objections 1 minor 23 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

T0 review · grok-4.3

Optimizing a few policy rollouts via gradient ascent on a learned latent world model raises returns and beats gradient-free MPC on continuous control tasks.

2026-05-25 06:42 UTC pith:IYYGEQYI

load-bearing objection Dream-MPC seeds a few policy trajectories then refines them with gradient ascent plus uncertainty penalty in a latent model, but the results do not isolate whether the gradients are what produces the edge over gradient-free MPC. the 2 major comments →

arxiv 2605.04568 v2 pith:IYYGEQYI submitted 2026-05-06 cs.LG cs.AIcs.RO

Dream-MPC: Gradient-Based Model Predictive Control with Latent Imagination

classification cs.LG cs.AIcs.RO
keywords model predictive controlmodel-based reinforcement learningworld modelsgradient-based optimizationlatent imaginationcontinuous controlpolicy improvement
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper sets out to demonstrate that gradient-based planning can succeed in model predictive control when trajectories are first sampled from a policy, then refined by ascent on the model's predicted returns, with added penalties for model uncertainty and reuse of prior optimizations across time steps. This hybrid avoids both the expense of population-based search and the limitations of a fixed policy by letting the model directly shape action sequences. A reader would care because many practical control problems involve high-dimensional continuous actions where pure gradient-free methods become slow, so a workable gradient route could make planning feasible at scale. The experiments across 24 tasks supply the evidence that the resulting controller improves the base policy and surpasses both gradient-free MPC and prior state-of-the-art agents.

Core claim

Dream-MPC generates a small number of candidate trajectories by rolling out the current policy, then performs gradient ascent on each trajectory inside the latent space of a learned world model; uncertainty regularization discourages the optimizer from exploiting model errors while amortization reuses previously optimized action sequences to reduce computation per step. When evaluated on 24 continuous control benchmarks the procedure measurably lifts the underlying policy's returns and exceeds the performance of gradient-free MPC as well as existing hybrid baselines.

What carries the argument

Gradient ascent on policy-initialized trajectories inside the latent dynamics model, regularized by uncertainty estimates and amortized across time steps.

Load-bearing premise

The world model's gradients must point toward higher true returns instead of letting the optimizer exploit inaccuracies in the model.

What would settle it

On the same 24 tasks, removing uncertainty regularization causes Dream-MPC to produce lower returns than the unoptimized policy or to fall behind gradient-free MPC.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • The base policy receives an immediate performance boost from the planning layer without retraining.
  • Gradient-based trajectory refinement can replace population-based search while remaining computationally lighter.
  • Amortization across time steps keeps the per-step cost of gradient optimization manageable.
  • Uncertainty penalties allow the method to stay robust even when the learned model is imperfect.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same latent model could support directed search in settings where sampling alone is inefficient, such as very high-dimensional action spaces.
  • Extending the amortization window or coupling it with longer-horizon policies might further reduce the number of gradient steps needed at runtime.
  • If the world model is replaced by one trained on real-world data rather than simulation, the gradient-based refinement could transfer more readily than pure policy or search methods.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 1 minor

Summary. The paper introduces Dream-MPC, a hybrid model-based RL method that generates candidate trajectories from a rolled-out policy and refines each via gradient ascent on a learned latent world model, using uncertainty regularization and temporal amortization of actions. It claims this yields significant policy improvement and outperforms both gradient-free MPC and state-of-the-art baselines across 24 continuous-control tasks.

Significance. If the reported gains are robust, the work offers a computationally lighter alternative to population-based planning while retaining the benefits of a policy prior; the public code release supports reproducibility and further testing of the gradient-based component.

major comments (2)
  1. [Abstract and §4] Abstract and §4 (Experiments): the central claim of outperformance on 24 tasks is stated without reference to error bars, statistical significance tests, baseline implementation details, or data-exclusion criteria. This information is load-bearing for attributing gains to the gradient-based optimizer rather than implementation variance.
  2. [§3 and §4.3] §3 (Method) and §4.3 (Ablations): the premise that world-model gradients reliably point toward higher return and that the uncertainty term keeps optimization inside the model's trustworthy region is invoked to justify replacing gradient-free search, yet no direct diagnostic (e.g., gradient-norm vs. true-return correlation or failure-case analysis) is supplied to verify the premise on the reported tasks.
minor comments (1)
  1. [§3.2] Notation for the uncertainty-regularized objective could be clarified with an explicit equation number when first introduced.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. The comments correctly identify areas where additional statistical details and diagnostics would strengthen the manuscript. We address each point below and will revise accordingly.

read point-by-point responses
  1. Referee: [Abstract and §4] Abstract and §4 (Experiments): the central claim of outperformance on 24 tasks is stated without reference to error bars, statistical significance tests, baseline implementation details, or data-exclusion criteria. This information is load-bearing for attributing gains to the gradient-based optimizer rather than implementation variance.

    Authors: We agree that explicit references to these elements are necessary. The experiments section already reports results over multiple random seeds with error bars in the figures, but we will add explicit mentions in the abstract and §4, include statistical significance tests (e.g., Wilcoxon signed-rank tests), provide full baseline implementation details and hyperparameters, and state that no data points were excluded. These changes will be incorporated in the revision. revision: yes

  2. Referee: [§3 and §4.3] §3 (Method) and §4.3 (Ablations): the premise that world-model gradients reliably point toward higher return and that the uncertainty term keeps optimization inside the model's trustworthy region is invoked to justify replacing gradient-free search, yet no direct diagnostic (e.g., gradient-norm vs. true-return correlation or failure-case analysis) is supplied to verify the premise on the reported tasks.

    Authors: We acknowledge the value of direct diagnostics to support the premise. While the ablations demonstrate the benefit of uncertainty regularization, we will add a new analysis (in §4.3 or an appendix) that includes gradient-norm vs. return correlation on selected tasks and examines failure cases where the uncertainty term prevents divergence. This will be included in the revised manuscript. revision: yes

Circularity Check

0 steps flagged

No circularity; empirical validation on external benchmarks

full rationale

The paper proposes Dream-MPC, a gradient-based MPC method using a learned world model, uncertainty regularization, and action amortization. Performance is assessed via direct comparison on 24 continuous control tasks against gradient-free MPC and baselines. No equations, fitted parameters renamed as predictions, self-definitional loops, or load-bearing self-citations appear in the derivation. The central claims rest on external task results rather than reducing to internal definitions or prior author work by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on the domain assumption that a learned world model yields usable gradients for return maximization; no free parameters or invented entities are named in the abstract.

axioms (1)
  • domain assumption A learned world model supplies gradients that improve expected return when followed from policy-generated seeds
    Invoked when the method switches from gradient-free to gradient-based trajectory optimization

pith-pipeline@v0.9.0 · 5716 in / 1221 out tokens · 19691 ms · 2026-05-25T06:42:40.029201+00:00 · methodology

0 comments
read the original abstract

State-of-the-art model-based Reinforcement Learning (RL) approaches either use gradient-free, population-based methods for planning, learned policy networks, or a combination of policy networks and planning. Hybrid approaches that combine Model Predictive Control (MPC) with a learned model and a policy prior to leverage the advantages of both paradigms have shown promising results. However, these approaches typically rely on gradient-free optimization methods, which can be computationally expensive for high-dimensional control tasks. While gradient-based methods are a promising alternative, recent works have empirically shown that gradient-based methods often perform worse than their gradient-free counterparts. We propose Dream-MPC, a novel approach that generates few candidate trajectories from a rolled-out policy and optimizes each trajectory by gradient ascent using a learned world model, uncertainty regularization and amortization of optimization iterations over time by reusing previously optimized actions. Our results on 24 continuous control tasks show that Dream-MPC can significantly improve the performance of the underlying policy and can outperform gradient-free MPC and state-of-the-art baselines. Code and videos are available at https://dream-mpc.github.io.

Figures

Figures reproduced from arXiv: 2605.04568 by Jonathan Spieler, Sven Behnke.

Figure 1
Figure 1. Figure 1: Aggregate performance metrics. Optimality gap, in￾terquartile median (IQM), mean and median normalized scores with 95% confidence intervals of different methods. Notably, Dream-MPC with a strong policy achieves the best results. ficiency and generalization, especially in complex and high￾dimensional environments (Byravan et al., 2022). Model￾based RL, on the other hand, can be more sample-efficient and can… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed approach. Dream-MPC optimizes action sequences rolled out from a policy network π in latent space z with gradient-based MPC. N candidate trajectories are sampled from the policy prior and optimized for I iterations using gradient ascent to maximize the objective J. The first action with the highest predicted return is applied, and the procedure is repeated for the next time step. T… view at source ↗
Figure 3
Figure 3. Figure 3: Performance profiles. Score distributions across all tasks, which provides insights into the variance of the performance. Detailed results are included in Tabs. 11 to 13. We first evaluate the performance of Dream-MPC when re￾placing the MPPI planner by our proposed gradient-based MPC planner at test time using (pre-)trained TD-MPC2 and BMPC models, respectively. For more details on the baselines refer to … view at source ↗
Figure 4
Figure 4. Figure 4: Learning curves for four tasks from the DeepMind Control Suite. The line represents the mean episodic return and the shaded area the 95% confidence interval across 3 seeds. trajectories. While combining gradient-based MPC with imitation learning is an interesting research direction, we leave this for future work view at source ↗
Figure 5
Figure 5. Figure 5: Ablations. (a) Performance of different Dream-MPC (TD-MPC2) variants demonstrating the importance of each design choice. (b) Performance of Dream-MPC (TD-MPC2) with different uncertainty regularization and action reuse coefficients. The line represents the mean episodic return and the shaded area the 95% confidence interval across 3 seeds. We further evaluate the performance of fully trained BMPC agents wi… view at source ↗
Figure 6
Figure 6. Figure 6: DeepMind Control Suite benchmarking domains (Tassa et al., 2020) view at source ↗
Figure 7
Figure 7. Figure 7: Meta-World manipulation tasks. We consider eight different tasks from the Meta-World Benchmark. We further consider following eight tasks from the twelve benchmarking locomotion tasks of HumanoidBench: • Balance Hard: Balance on the unstable board while the spherical pivot beneath the board does move, • Balance Simple: Balance on the unstable board while the spherical pivot beneath the board does not move,… view at source ↗
Figure 8
Figure 8. Figure 8: HumanoidBench locomotion tasks. We consider eight tasks from the HumanoidBench locomotion benchmark that cover a wide variety of interactions and difficulties. This figure illustrates an initial state for each task view at source ↗
Figure 9
Figure 9. Figure 9: Parameter sweep. Performance of trained BMPC agents with Dream-MPC at test time when varying the number of candidates, horizon and number of optimization iterations. When varying one hyperparameter, the others are fixed to their default value. We also include the performance of the learned policy πθ and the default values of one iteration, a horizon of three and five candidate trajectories. 18 view at source ↗
Figure 9
Figure 9. Figure 9: Learning curves for the DeepMind Control Suite. The line represents the mean episodic return and the shaded area the 95% confidence interval across 3 seeds. 0 250K 500K 750K 1M 0.0 0.5 1.0 Episode success Assembly 0 250K 500K 750K 1M 0.0 0.5 1.0 Button Press 0 250K 500K 750K 1M 0.0 0.5 1.0 Disassemble 0 250K 500K 750K 1M 0.0 0.5 1.0 Lever Pull 0 250K 500K 750K 1M Environment steps 0.0 0.5 1.0 Episode succe… view at source ↗
Figure 10
Figure 10. Figure 10: Learning curves for the DeepMind Control Suite. The line represents the mean episodic return and the shaded area the 95% confidence interval across 3 seeds. 0 250K 500K 750K 1M 0.0 0.5 1.0 Episode success Assembly 0 250K 500K 750K 1M 0.0 0.5 1.0 Button Press 0 250K 500K 750K 1M 0.0 0.5 1.0 Disassemble 0 250K 500K 750K 1M 0.0 0.5 1.0 Lever Pull 0 250K 500K 750K 1M Environment steps 0.0 0.5 1.0 Episode succ… view at source ↗
Figure 11
Figure 11. Figure 11: Learning curves for Meta-World. The line represents the mean episodic return and the shaded area the 95% confidence interval across 3 seeds. 0 500K 1M 1.5M 2M 0 3000 6000 9000 12000 Episode return Reach 0 500K 1M 1.5M 2M 0 250 500 750 1000 Hurdle 0 500K 1M 1.5M 2M 0 250 500 750 1000 Maze 0 500K 1M 1.5M 2M 0 250 500 750 1000 Run 0 500K 1M 1.5M 2M Environment steps 0 250 500 750 1000 Episode return Balance … view at source ↗
Figure 12
Figure 12. Figure 12: Learning curves for HumanoidBench. The line represents the mean episodic return and the shaded area the 95% confidence interval across 3 seeds. 19 view at source ↗
Figure 12
Figure 12. Figure 12: shows the performance of trained BMPC agents with Dream-MPC at test time when varying the number of candidates, horizon and number of optimization iterations. When varying one hyperparameter, the others are fixed to their default value. 1 3 5 7 9 550 600 650 Episode return 1 3 5 7 9 Acrobot Swingup 1 3 5 7 9 1 3 5 7 9 300 400 500 Episode return 1 3 5 7 9 Humanoid Run 1 3 5 7 9 1 3 5 7 9 Candidates 200 400… view at source ↗
Figure 13
Figure 13. Figure 13: Learning curves for four tasks from the DeepMind Control Suite. The line represents the mean episodic return and the shaded area the 95% confidence interval across 3 seeds. are performed with only RGB visual observations with a resolution of 64 × 64. We evaluate the performance of our method when enabling planning already during training. The learning curves are shown in view at source ↗
Figure 14
Figure 14. Figure 14: Planner gradients of Grad-MPC and Dream-MPC. For different planning horizons on the Pendulum-v1 environment using the ground truth (simulator) and learned dynamics model respectively and state observations. The values are represented by their mean and standard deviation for three different random seeds. The default hyperparameters provided in Tab. 18 are used unless otherwise specified. 0 51K 102K 0.0 1.5… view at source ↗
Figure 15
Figure 15. Figure 15: Expected Signal-to-Noise Ratio (ESNR) of the planner gradients of Grad-MPC and Dream-MPC. Calculated via Eq. (17) for different planning horizons on the Pendulum-v1 environment using the ground truth (simulator) and learned dynamics model respectively and state observations. The values are represented by their mean and standard deviation for three different random seeds. The default hyperparameters provid… view at source ↗
Figure 15
Figure 15. Figure 15: Expected Signal-to-Noise Ratio (ESNR) of the planner gradients of Grad-MPC and Dream-MPC. Calculated via Eq. (17) for different planning horizons on the Pendulum-v1 environment using the ground truth (simulator) and learned dynamics model respectively and state observations. The values are represented by their mean and standard deviation for three different random seeds. The default hyperparameters provid… view at source ↗
Figure 16
Figure 16. Figure 16: Analysis of predicted returns over the number of environment steps for Acrobot Swingup. E.4. Implementation Details We use PyTorch (Paszke et al., 2019) implementations of SAC+AE2 , PlaNet and Dreamer3 that are distributed under MIT license and also base the implementations of Grad-MPC and of our method on the latter. The hyperparameters are listed in Tab. 18. We use the default hyperparameters for SAC+AE… view at source ↗
Figure 16
Figure 16. Figure 16: Analysis of predicted returns over the number of environment steps for Acrobot Swingup. E.4. Implementation Details We use PyTorch (Paszke et al., 2019) implementations of SAC+AE2 , PlaNet and Dreamer3 that are distributed under MIT license and also base the implementations of Grad-MPC and of our method on the latter. The hyperparameters are listed in Tab. 18. We use the default hyperparameters for SAC+AE… view at source ↗
Figure 17
Figure 17. Figure 17: Value estimations for Dog Run. The lines represent the mean values and the shaded areas the 95% confidence interval across 3 seeds. Similar to Lin et al. (2026), we find that for high-dimensional problems, such as Dog Run, TD-MPC2 tends to overestimate the values, although the gap is in our case smaller and also changes from an overestimation in the beginning to an underestimation at the end of the traini… view at source ↗
Figure 18
Figure 18. Figure 18: Effects of value approximation errors. Analysis for BMPC on Dog Run. (a) Absolute value error vs. episode return. (b) Mean variance of predicted Q-values vs. episode return. (c) Mean variance of predicted Q-values vs. absolute value error. 31 [PITH_FULL_IMAGE:figures/full_fig_p031_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Value estimation comparison for BMPC policy on Dog Run. 32 [PITH_FULL_IMAGE:figures/full_fig_p032_19.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages · 4 internal anchors

  1. [1]

    Combating the Compounding-Error Problem with a Multi-step Model

    URL http://arxiv.org/abs/ 1905.13320. Bharadhwaj, H., Xie, K., and Shkurti, F. Model-predictive control via cross-entropy and gradient-based optimization. In2nd Annual Conference on Learning for Dynamics and Control (L4DC),

  2. [2]

    World Models

    URL http://arxiv.org/abs/1803.10122. Haarnoja, T., Zhou, A., Abbeel, P., and Levine, S. Soft Actor-Critic: Off-policy maximum entropy deep rein- forcement learning with a stochastic actor. In35th In- ternational Conference on Machine Learning (ICML),

  3. [3]

    Model-Based Planning with Discrete and Continuous Actions

    URL http://arxiv.org/abs/1705.07177. 9 Dream-MPC: Gradient-Based Model Predictive Control with Latent Imagination Hubert, T., Schrittwieser, J., Antonoglou, I., Barekatain, M., Schmitt, S., and Silver, D. Learning and planning in complex action spaces. In38th International Conference on Machine Learning (ICML),

  4. [4]

    A path towards autonomous machine intelli- gence version 0.9.2, 2022-06-27,

    LeCun, Y . A path towards autonomous machine intelli- gence version 0.9.2, 2022-06-27,

  5. [5]

    Gradient-based planning with world models.arXiv:2312.17227,

    URL http: //arxiv.org/abs/2312.17227. Seo, Y ., Sferrazza, C., Chen, J., Shi, G., Duan, R., and Abbeel, P. Learning sim-to-real humanoid locomotion in 15 minutes

  6. [6]

    Learning sim-to-real humanoid locomotion in 15 minutes, 2025a

    URL https://arxiv.org/ abs/2512.01996. Sferrazza, C., Huang, D.-M., Lin, X., Lee, Y ., and Abbeel, P. HumanoidBench: Simulated humanoid benchmark for whole-body locomotion and manipulation. InRobotics: Science and Systems Confererence (RSS),

  7. [7]

    Model Predictive Path Integral Control using Covariance Variable Importance Sampling

    URL http://arxiv. org/abs/1509.01149. Williams, G., Wagener, N., Goldfain, B., Drews, P., Rehg, J. M., Boots, B., and Theodorou, E. A. Information the- oretic MPC for model-based reinforcement learning. In IEEE International Conference on Robotics and Automa- tion (ICRA),

  8. [8]

    org/abs/1912.11206

    URL https://arxiv. org/abs/1912.11206. Xie, K., Bharadhwaj, H., Hafner, D., Garg, A., and Shkurti, F. Latent skill planning for exploration and transfer. In9th International Conference on Learning Representations (ICLR),

  9. [9]

    11 Dream-MPC: Gradient-Based Model Predictive Control with Latent Imagination A. Limitations and Future Work Fixed optimization parameters.Our experiments suggest that it may be beneficial to dynamically adapt the optimization parameters such as the action optimization step size and number of iterations to further improve the performance, especially for h...

  10. [10]

    Please refer to Yu et al

    and action space (dim(A) = 4). Please refer to Yu et al. (2019) for the definitions of the reward functions and success metrics used in the Meta-World tasks. Assembly Button Press Disassemble Lever Pull Pick Place Wall Push Back Shelf Place Window Open Figure 7.Meta-World manipulation tasks.We consider eight different tasks from the Meta-World Benchmark. ...

  11. [11]

    For the experiments with (pre-)trained models, we use the models provided by Hansen et al

    Details on TD-MPC2 can be found in Section C.1. For the experiments with (pre-)trained models, we use the models provided by Hansen et al. (2024) for the DeepMind Control Suite and Meta-World, except for Cartpole Swingup Sparse, Dog Run, Dog Walk, Humanoid Run and Humanoid Walk because some checkpoints cannot be loaded after code restructuring1. Thus, we ...

  12. [12]

    Policy-guided MPC.TD-MPC2 uses Model Predictive Path Integral (MPPI) (Williams et al., 2015

    Since we only perform single-task experiments in this work, all models contain around 5M parameters for TD-MPC2. Policy-guided MPC.TD-MPC2 uses Model Predictive Path Integral (MPPI) (Williams et al., 2015

  13. [13]

    MPPI iteratively samples action sequences (at, at+1,

    for local trajectory optimization, which is a gradient-free, sampling-based MPC method. MPPI iteratively samples action sequences (at, at+1, . . . , at+H ) of length H from N(µ, σ 2), evaluates their expected return by rolling out latent trajectories with the model, and updates the parameters µ, σ of a time-dependent multivariate Gaussian with diagonal co...

  14. [14]

    While Dream-MPC can improve the performance of the policy for TD-MPC2, it cannot consistently match the performance of MPPI

    We find that having a good policy is important because it leads to better value estimates, which are crucial for gradient-based MPC. While Dream-MPC can improve the performance of the policy for TD-MPC2, it cannot consistently match the performance of MPPI. Since the performance of the policy is quite weak as shown in Tabs. 14 to 16, this fact favours MPP...

  15. [15]

    t+HX n=τ rn # ,(12) V k N (sτ ) =E qθ,πϕ

    to show that it also works with other model-based RL algorithms. Dreamer learns a latent dynamics model, often referred to as a world model, consisting of the following components: • Representation model:p θ(st|st−1, at−1, ot) • Transition model:q θ(st|st−1, at−1) • Reward model:q θ(rt|st) • Observation model (only used as an additional learning signal):q...

  16. [16]

    Initialize model parametersθ, ϕ, ψrandomly

    Algorithm 2Dream-MPC integration into Dreamer Input:Representation model pθ(st|st−1, at−1, ot), transition model qθ(st|st−1, at−1), reward model qθ(rt|st), value function model vψ(st), policy model πϕ(at|st), exploration noise p(ϵ), action repeat R, seed episodes S, collect interval C, batch size B, chunk lengthL, learning rateη Initialize datasetDwithSra...

  17. [17]

    algorithm for image-based observations and the Policy+Grad-MPC method proposed in (S V et al., 2023). Note that Policy+Grad-MPC and Dream-MPC both share the general idea of using a policy 0 250K 500K 750K 1M Environment steps 0 80 160 240Episode return Acrobot Swingup 0 250K 500K 750K 1M Environment steps 0 250 500 750 Cheetah Run 0 250K 500K 750K 1M Envi...

  18. [18]

    In contrast to PlaNet (CEM) and Grad-MPC, which both use 1000×10×12 =120 000 evaluations of the world model at each time step, our method only requires 5×1×15 = 75 evaluations

    We find that our method can not only outperform the baselines, but also that planning during training can improve the sample efficiency without leading to premature convergence. In contrast to PlaNet (CEM) and Grad-MPC, which both use 1000×10×12 =120 000 evaluations of the world model at each time step, our method only requires 5×1×15 = 75 evaluations. Th...

  19. [19]

    suggest that learned models can improve ESNR compared to using the ground truth dynamics for some problems, indicating the possibility of further improvement. While the ESNR significantly suffers for horizons greater than ten for Grad-MPC using the learned dynamics model, the ESNR for Dream-MPC remains much more stable for increasing horizons. Together wi...

  20. [20]

    (2021), except for the action repeat, which we set to two for a fair comparison

    We use the default hyperparameters for SAC+AE as described in Yarats et al. (2021), except for the action repeat, which we set to two for a fair comparison. — Appendices continue on next page — 2https://github.com/denisyarats/pytorch_sac_ae 3https://github.com/yusukeurakami/dreamer-pytorch 27 Dream-MPC: Gradient-Based Model Predictive Control with Latent ...

  21. [21]

    Max. episode length 1000 Action repeat 2 Experience size 1000000 Embedding size 1024 Hidden size 200 Belief size 200 State size 30 Exploration noise 0.3 Seed episodes 5 Collect interval 100 Batch size 50 Overshooting distance 0 Overshooting KL beta 0 Overshooting reward scale 0 Global KL beta 0 Free nats 3 Bit depth 5 Dreamer & Dream-MPC Planning horizon ...

  22. [22]

    provides only limited experimental results and lacks in-depth implementation details. While it shows that gradient-based MPC with a policy network is promising for two sparse-reward tasks from the DeepMind Control Suite, it does not provide a full evaluation of the method in diverse settings such as different benchmarks, different world models or types of...

  23. [23]

    For BMPC, we use five Q-networks in all following experiments instead of two V-networks to isolate potential confounding effects arising from using V-networks instead. 0 250K 500K 750K 1M Environment steps 0 50 100 150 200Value TD-MPC2 0 250K 500K 750K 1M Environment steps 0 50 100 150 200 TD-MPC2 w/o MPC 0 250K 500K 750K 1M Environment steps 0 50 100 150...