Pith.

Robotics · arXiv:2603.25685 v2 · submitted 26 Mar 2026

Persistent Robot World Models: Stabilizing Multi-Step Rollouts via Reinforcement Learning

Jai Bardhan, Patrik Drozdik, Josef Sivic, Vladimir Petrik

The paper says

“Fourth, we show that our approach establishes a new state-of-the-art for rollout fidelity on the DROID dataset, outperforming the strongest baseline on all metrics (e.g., LPIPS reduced by 14% on external cameras, SSIM improved by 9.1% on the wrist camera), winning 98% of paired comparisons, and achieving an 80% preference rate in a blind human study.”

from the abstract · picked out as the main claim by a classifier

Pith says

Pith has not reviewed this paper.

Get a Pith Review, $5

Reviews

Reviews by people

No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Action-conditioned robot world models generate future video frames of the manipulated scene given a robot action sequence, offering a promising alternative for simulating tasks that are difficult to model with traditional physics engines. However, these models are optimized for short-term prediction and break down when deployed autoregressively: each predicted clip feeds back as context for the next, causing errors to compound and visual quality to rapidly degrade. We address this through the following contributions. First, we introduce a reinforcement learning (RL) post-training scheme that trains the world model on its own autoregressive rollouts rather than on ground-truth histories. We achieve this by adapting a recent contrastive RL objective for diffusion models to our setting and show that its convergence guarantees carry over exactly. Second, we design a training protocol that generates and compares multiple candidate variable-length futures from the same rollout state, reinforcing higher-fidelity predictions over lower-fidelity ones. Third, we develop efficient, multi-view visual fidelity rewards that combine complementary perceptual metrics across camera views and are aggregated at the clip level for dense, low-variance training signal. Fourth, we show that our approach establishes a new state-of-the-art for rollout fidelity on the DROID dataset, outperforming the strongest baseline on all metrics (e.g., LPIPS reduced by 14% on external cameras, SSIM improved by 9.1% on the wrist camera), winning 98% of paired comparisons, and achieving an 80% preference rate in a blind human study.

The paper's own sentences, arranged. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-09-23 · no language-model review

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.09 · states limitation 0.13 · human participants 0.68 · controlled comparison 0.94 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.97 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v2). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Action-conditioned robot world models generate future video frames of the manipulated scene given a robot action sequence, offering a promising alternative for simulating tasks that are difficult to model with traditional physics engines. background 1.00 0.06 0.03 asserted 0.66
However, these models are optimized for short-term prediction and break down when deployed autoregressively: each predicted clip feeds back as context for the next, causing errors to compound and visual quality to rapidly degrade. background 0.66 0.13 0.04 asserted 1.00
We address this through the following contributions. contribution 0.97 0.40 0.02 unclear 0.75
First, we introduce a reinforcement learning (RL) post-training scheme that trains the world model on its own autoregressive rollouts rather than on ground-truth histories. contribution 0.96 0.94 0.07 asserted 0.99
We achieve this by adapting a recent contrastive RL objective for diffusion models to our setting and show that its convergence guarantees carry over exactly. method 0.77 0.83 0.04 asserted 1.00
Second, we design a training protocol that generates and compares multiple candidate variable-length futures from the same rollout state, reinforcing higher-fidelity predictions over lower-fidelity ones. contribution 0.89 0.96 0.04 asserted 0.98
Third, we develop efficient, multi-view visual fidelity rewards that combine complementary perceptual metrics across camera views and are aggregated at the clip level for dense, low-variance training signal. contribution 0.81 0.95 0.04 asserted 0.99
Fourth, we show that our approach establishes a new state-of-the-art for rollout fidelity on the DROID dataset, outperforming the strongest baseline on all metrics (e.g., LPIPS reduced by 14% on external cameras, SSIM improved by 9.1% on the wrist camera), winning 98% of paired comparisons, and achieving an 80% preference rate in a blind human study. result 0.97 0.97 0.73 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.92

What was examined, and what was not

Sources read

  • arXiv listing abstract v2

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 6,437 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Cited by 5 Pith papers
How the literature uses it

5 arXiv papers cite this in 12 sentences

Each citing sentence is taken from the citing paper's LaTeX source and typed by Pith's classifier. A label appears at probability 0.8 or more (0.95 for support), a contest at 0.7 or more; the quoted words are the citing authors'. How each label was checked · JSON

Papers citing it · 7 sentences without a confident label

Compares with it 2 papers
Cites it as background 2 papers

What it cites · 15 sentences without a confident label

Builds on 2 papers
  • We then formulate post-training as an online reinforcement learning (RL) problem and derive a tractable objective by adapting a contrastive forward-process training designed for velocity-prediction flow-matching models [ref] — to the EDM [ref] $x_0$-prediction parameterization of Ctrl-World's SVD backbone; we show that the branch construction and policy-improvement guarantees carry over exactly, with no $σ$-dependent correction terms (Sec. [ref]).

    DiffusionNFT: Online Diffusion Reinforcement with Forward Process · Improving Robot World Models with Reinforcement Learning · 0.85

  • We then formulate post-training as an online reinforcement learning (RL) problem and derive a tractable objective by adapting a contrastive forward-process training designed for velocity-prediction flow-matching models [ref] — to the EDM [ref] $x_0$-prediction parameterization of Ctrl-World's SVD backbone; we show that the branch construction and policy-improvement guarantees carry over exactly, with no $σ$-dependent correction terms (Sec. [ref]).

    Elucidating the Design Space of Diffusion-Based Generative Models · Improving Robot World Models with Reinforcement Learning · 0.85

Uses the method or tool of 4 papers
Uses data or values from 1 paper
Compares with 3 papers
  • Our model establishes a new state-of-the-art for autoregressive rollout quality on the DROID dataset, consistently outperforming all baselines—WPE [ref] , IRASim [ref] , and Ctrl-World [ref] —across every metric (Table [ref]).

    Ctrl-World: A Controllable Generative World Model for Robot Manipulation · Experiments · 1.00

  • Our model establishes a new state-of-the-art for autoregressive rollout quality on the DROID dataset, consistently outperforming all baselines—WPE [ref] , IRASim [ref] , and Ctrl-World [ref] —across every metric (Table [ref]).

    2506.00613 · Experiments · 1.00

  • DPPO [ref] applies RL fine-tuning to diffusion-based robot action policies; our work instead post-trains the world model itself.

    Diffusion Policy Policy Optimization · Related work · 0.88

Cites as support 1 paper
  • RLVR-World [ref] applies RL with verifiable rewards to improve world model transition quality, further evidencing that RL objectives outperform MLE for rollout fidelity.

    2505.13934 · Related work · 1.00

Cites as background 16 papers

Challenges and discussion

Know something about this paper? Leave a comment.

No comments yet.

Sign in to comment