PAPER REPORTENAll readings ↗

Learning Massively Multitask World Models for Continuous Control

English reading report: Method, equations, original figures, experiments and reproducibility.

Authors: Nicklas Hansen; Hao Su; Xiaolong Wang

Affiliations: University of California San Diego

Source: 2511.19584 ↗ · Catalog record

Reading: 319 / 558 · 6 original figures & tables · ~20 min ·

1. Paper overview

In one sentence: A demonstration-pretrained latent world model improves one controller across 200 tasks, while trailing specialists and showing uneven performance without feedback. e02e04e07e11e15e17e18

At a glanceWhat to know
Research problem
Source description

Can one agent improve through online RL across 200 tasks and ten domains despite incompatible embodiments, reward scales and horizons? MMBench supplies instructions, demonstrations and optional images, with 20 additional held-out tasks. Its score combines domain-specific success criteria and normalized returns; 0.44 is not a universal 44% success rate. e02e03

Core mechanism
Source description

MMBench combines 159 existing tasks, 22 new variants in existing domains and 19 MiniArcade training tasks. The full MiniArcade suite has 22 tasks, so release size differs from the training subset. e02e03

A key reported resultMMBench state-based multitask control: 0.438; Table 9 rounds this to 0.44.

Average normalized score, higher is better. 200 training tasks; default Newt; 100M total online steps plus demonstrations.

Figure 7 and Section 4.1 show higher overall performance than multitask PPO, FastTD3 and TD-MPC2. Specialist populations score 0.65 at 100M total steps and 0.80 at Table 9’s labeled 1B budget. Shared control trails specialists. Equal online steps do not equalize demonstration-generation cost or total model capacity. e11e15e18e03

Reading caution
Source description

Training largely uses privileged state inputs in simulation. Atari uses raw RAM and shortened episodes, limiting transfer. One fixed instruction per training task restricts language diversity. e03e20

Core contributions

  • Source description

    MMBench combines 159 existing tasks, 22 new variants in existing domains and 19 MiniArcade training tasks. The full MiniArcade suite has 22 tasks, so release size differs from the training subset. e02e03

  • Source description

    Newt combines decoder-free prediction, language conditioning, model-based pretraining and persistent action supervision, supported by asynchronous interaction and distributed updates. e04e05e06e07

  • Reader analysis

    The advance is shared multitask learning and adaptation, rather than superiority to specialists or demonstrated physical robot deployment. e11e13e18

Figure 6. A shared latent model connects task instructions to action-conditioned planning. Original paper, p. 5 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read the left side as an interaction loop across MMBench environments. CLIP encodes the embodiment and task description, DINOv2 supplies optional image features, and state observations enter the shared Newt model. The hatched action dimensions are invalid for that task. The right inset expands the learning mechanism: the initial observation is encoded, latent states propagate forward, and future observation encodings provide targets. Red stop-gradient marks belong to those future target branches, consistent with Equation (2). Equation (1) supplies the action and language conditioning of dynamics that the compact inset leaves implicit; its a, r and q heads should not be read as a complete planner diagram. e02e04e05e06e09e11

What it supports. The model learns representations useful for predicting controlled outcomes. Its rewards, values and policy prior support trajectory search, while the self-prediction objective anchors latent transitions to observed successors. The central action columns represent a plan; Appendix E explains why normal execution takes only the first action before observing again.

Where the evidence stops. The diagram shows training targets as well as runtime inputs. Future observations are not available to the online planner, and the inset is not a future-image decoder. Optional RGB does not make the reported main benchmark a camera-only experiment.

2. Motivation

2.1 The problem and the proposed response

Source description

Can one agent improve through online RL across 200 tasks and ten domains despite incompatible embodiments, reward scales and horizons? MMBench supplies instructions, demonstrations and optional images, with 20 additional held-out tasks. Its score combines domain-specific success criteria and normalized returns; 0.44 is not a universal 44% success rate. e02e03

2.2 What this reading follows

Newt asks how far a single controller can go when it learns from both demonstrations and continued interaction across many environments. Its world model predicts latent consequences, rewards and values, then uses those predictions to search for actions. Language distinguishes the intended task; a pretrained policy prior makes exploration and early planning more tractable. The figures below connect this mechanism to the evidence: aggregate improvements over multitask baselines, uneven domain behavior, a persistent gap to specialists and selective open-loop competence. The visual-input experiment adds another qualification: useful manipulation gains barely change the overall score, and the comparison includes extra finetuning. e02e04e07e11e15e17e18

3. Research context

We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.

Catalog dimensionRecorded classification
Major categoryNot assigned
ArchitectureNot assigned
Prediction paradigmNot assigned
QuadrantNot assigned

This table preserves the labels recorded at reading time. The current major category is WAMs. View the current classification.

3.1 Evidence-based assessment

Insufficient evidence to decide

Reader analysis

The catalog is unassigned, so no substantive label can be supported or contradicted. Architecture evidence supports action-conditioned latent dynamics used by an explicit planner, with distinct policy-prior and prediction/value MLPs. This is neither inverse dynamics nor joint future-video/action generation. One multitask agent alone does not establish a “One Model” classification. e04e05e06e09

This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.

4. Problem formulation

4.1 Inputs and outputs

InputsOutputs
  • Masked 128-dimensional state vectors and embodiment/task instructions
  • Optional 224×224 RGB observations
  • Replay subsequences of observations, actions, rewards and successor observations
  • Predicted successor latents, rewards and action values
  • A stochastic policy prior and a planned continuous action sequence; normally only its first action is executed

4.2 Equations and their role

L(θ)=EτB[t=0Hλt(ztsg ⁣(h(sstate,t,xt,g))22+CE(r^t,rt)+CE(q^t,qt))]\mathcal{L}(\theta)=\mathbb{E}_{\tau\sim\mathcal{B}}\left[\sum_{t=0}^{H}\lambda^t\left(\left\|z_t^{\prime}-\operatorname{sg}\!\left(h(s_{\mathrm{state},t}^{\prime},x_t^{\prime},g)\right)\right\|_2^2+\ell_{\mathrm{CE}}(\hat r_t,r_t)+\ell_{\mathrm{CE}}(\hat q_t,q_t)\right)\right]
Equation (2): θ denotes parameters; τ is a subsequence from replay B, H the rollout horizon and λ temporal weighting. z′ is the predicted successor latent; h encodes successor state s′state, optional image embedding x′ and language embedding g. sg stops target gradients. Cross-entropy ℓCE compares predicted reward/value with reward r and one-step TD target q. e04e05
Lp(θ)=EτB[t=0Hλt(p(zt,g)at22Q(zt,p(zt,g),g)H(p(zt,g)))]\mathcal{L}_p(\theta)=\mathbb{E}_{\tau\sim\mathcal{B}}\left[\sum_{t=0}^{H}\lambda^t\left(\|p(z_t,g)-a_t\|_2^2-Q(z_t,p(z_t,g),g)-\mathcal{H}(p(\cdot\mid z_t,g))\right)\right]
Equation (3): prior p learns from replay actions a, action value Q and entropy ℋ. Starting from the encoded initial observation, replay actions advance latent states. BC therefore supervises latent rollouts and distills previously planned actions. Pretraining disables the negative-Q term; actual weights are specified in Table 7. e06e07e10

5. Method in detail

5.1 Learn consequences in a space shaped by control

Source description

Start with an observed state and a description of the embodiment and goal. Frozen text features help the encoder distinguish tasks even when their physical observations overlap, as in RoboDesk. The resulting latent state is the starting point for predicting what an action will do. Newt trains these predictions against encodings of actual successor observations, stopping gradients through the target branch. Reward prediction and TD-value learning add supervision relevant to choosing actions; reconstructing every image pixel is unnecessary for this objective. The model nevertheless has several distinct components: an encoder, dynamics, reward predictor, value ensemble and stochastic prior. Their joint optimization makes the shared representation useful for control, while the explicit search procedure converts predicted consequences into behavior. A single task-conditioned agent should therefore not be confused with a monolithic joint video-and-action generator. e02e04e05e06e09

5.2 Bridge a good imitation policy to an initially imperfect planner

Source description

A pretrained policy can already imitate useful actions while the value function remains inaccurate. Replacing that policy abruptly with unrestricted planning can reduce performance, because search exploits estimates that have not yet become reliable. Newt addresses this transition in several connected ways. It pretrains every learnable component on demonstrations, but temporarily removes Q maximization from the prior’s objective. At the start of online learning, planning is biased toward the pretrained action distribution; Appendix E specifies a bias schedule annealed from 2M to 12M steps. Demonstration subsequences continue to occupy half the training samples. Finally, the BC term remains active on latent rollouts, supervising both demonstrated actions and actions collected through planning. These mechanisms help maintain usable behavior while online experience improves the model and supplies opportunities beyond the initial demonstrations. e06e07e08e09

5.3 Separate a useful model from broad control claims

Reader analysis

Reader analysis: the strongest interpretation comes from keeping three comparisons separate. Figure 7 asks whether one multitask algorithm improves over other multitask algorithms. Table 9 asks whether sharing one agent beats separately trained specialists; its answer remains negative at the reported aggregate budgets. Table 8 asks whether useful behavior survives removal of feedback, and its answer depends strongly on the task. Executing a long sequence successfully in selected environments is evidence of some useful predictive structure, but the large retained-performance losses prevent treating it as general long-horizon reliability. Transfer adds encouraging evidence for initialization, yet Figure 9 and its accompanying prose disagree on the scratch endpoint, so the exact improvement remains unresolved. These boundaries motivate controlled follow-ups that hold checkpoints, starting states and training budgets fixed before assigning a gain to planning or visual information. e11e13e15e17e18

5.4 Training and inference

During training

Source description

There are 4,020 demonstrations, 10–40 per task, from specialist TD-MPC2 agents. Episodes failing success criteria or returning more than 25% below the task median are rejected. Appendix C specifies 5M expert-training steps, except 10M for Atari; the main baseline description omits this exception. e08e11

Source description

Pretrain all learnable components for 200,000 iterations, disabling only the policy objective’s Q-maximization term. Online updates sample demonstration/interaction buffers 50:50 and retain BC on replay actions. Initial planning favors the pretrained policy; Appendix E specifies annealing from 2M to 12M steps. e06e07e09

Source description

Defaults are approximately 20M learnable parameters, 100M total online steps, 10M replay capacity, batch 1,024 and update-to-data ratio 0.075. Adam rates are 3×10⁻⁴, or 1×10⁻⁴ for the encoder. Table 7 gives loss coefficients omitted from the schematic equations, including BC weight 10. e10e11

During inference

Source description

Closed-loop control replans after each observation using a shifted previous search distribution. Open-loop evaluation executes the full sequence without feedback or replanning. The prior proposes actions; model-based search selects the behavior executed in the simulator. e09

5.5 Implementation flow

  1. Encode the task and observation

    Frozen CLIP-ViT/B text features and optional frozen DINOv2/B image features have 512 and 768 dimensions. Concatenating these with state inputs feeds an MLP encoder. The default latent has 512 dimensions, normalized in simplices of dimension eight. No frame stacking is used. e02e04

  2. Predict action-conditioned consequences

    Separate MLPs implement dynamics d, reward R, value ensemble Q and policy prior p, conditioned on language. Dynamics predict embeddings rather than pixels. Reward/value learning uses log-space cross-entropy regression; TD targets use an EMA Q ensemble and a random-subset minimum. e04e05

  3. Search for actions

    A CEM/MPPI-style search fits a diagonal Gaussian over action sequences to valuable latent rollouts, including policy-prior candidates. Defaults are horizon 3, six iterations, population 512, 24 policy samples and 64 elites. e09

6. Experiments & results

Newt extends TD-MPC2 into a language-conditioned agent that learns latent dynamics from demonstrations and online interaction across MMBench. Its practical recipe pretrains the world model and policy prior, retains action supervision, and uses the model to plan. It improves over the evaluated multitask baselines, but specialist policies remain stronger and long open-loop execution is uneven. The evidence concerns simulated, predominantly state-based control.

6.1 Read the original evidence

Figure 7. The aggregate advantage is concentrated in particular domains. Original paper, p. 7 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Follow the thick blue Newt curve within each panel, comparing it with the green FastTD3, red TD-MPC2 and purple PPO curves. Dashed horizontal lines are BC baselines, with the brown line representing separate single-task BC policies. The horizontal axis counts total interaction across all training tasks, not a full budget for each individual task. Vertical scales differ between panels. Appendix A also defines different underlying scores: some domains use task success, while others rescale episode returns. Read domain rankings before forming an overall judgment, because a shared numeric range does not make every domain an identical evaluation problem. e03e11

What it supports. Newt’s clear strengths include DMControl, its extended suite and ManiSkill, while FastTD3 is stronger in parts of Meta-World and RoboDesk. The source’s overall multitask advantage therefore coexists with important local exceptions. Low Atari and MuJoCo curves further limit any claim that a single recipe solves the entire benchmark.

Where the evidence stops. These comparisons do not equalize access to demonstrations or every architecture choice. Newt’s multitask TD-MPC2 baseline lacks several additions at once, so this plot cannot isolate the causal contribution of any one change. Seed counts and uncertainty are not specified here.

Table 9. One shared agent remains behind a population of specialists. Original paper, p. 35 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Begin with the grouped column headers. Newt’s two columns describe a single shared agent at total interaction budgets of 20M and 100M steps. The three TD-MPC2 columns aggregate separate specialist agents, with the budgets summed across the population. Compare the two 100M columns for the equal-online-step comparison, then examine individual domain rows to see how the gap changes. The final 1B specialist column is the source’s labeled convergence budget, not a matched-budget baseline. The Total row gives the reported overall normalized score; the adjacent task-count column prevents interpreting each domain as an equally sized collection. e03e08e11e18

What it supports. At 100M total online steps, Newt scores 0.44 against the specialist population’s 0.65. The labeled 1B specialist score is 0.80. Thus beating the multitask baselines in Figure 7 does not establish superiority over specialist learning, even when the comparison uses the same total online interaction budget.

Where the evidence stops. Equal online budgets exclude the cost of Newt’s expert-generated demonstrations and do not match total model capacity. The caption uses 5M steps per expert; Appendix C separately specifies 10M for Atari. The source’s budget labels are retained without reconciliation.

Table 8. Long plans expose strong task dependence and substantial feedback dependence. Original paper, p. 33 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read each value as a fraction of the corresponding closed-loop control performance, rather than as raw return or success probability. The upper block evaluates four tasks after 16, 32 and 48 planned steps; the lower block uses 8, 16 and 24 steps for four different tasks. These time points correspond to roughly one-third, two-thirds and all of the plan. Read each row across to track deterioration, then compare the bold averages within its own block. Both horizons exceed the training horizon of three, but the two blocks use different tasks, so their means do not isolate horizon length by themselves. e09e17

What it supports. The 48-step group retains only 0.24 on average at its final point; the 24-step group retains 0.60. Pick Screwdriver and Push Green retain 0.83, showing useful open-loop capability in particular settings. Walker Walk at 0.25 and Assault at 0.08 show why feedback remains important.

Where the evidence stops. The main text’s claim of closely matching closed-loop performance in most tasks is broader than these numbers support. No uncertainty or absolute closed-loop denominators are reported in this table, and eight selected tasks do not establish open-loop capability across all 200 training tasks.

Table 2. Small aggregate visual gains conceal domain-level improvements and regressions. Original paper, p. 9 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. The first two rows compare the state agent’s overall normalized score with the score after adding RGB observations and finetuning for another 30M environment steps. Indented rows then show selected domains, not an exhaustive decomposition of the overall average. The Score column gives the resulting domain score; Gain compares that domain with its own state-based reference, not with the overall 0.438 row. Positive and negative colors make the tradeoffs visible. Section 4.1 supplies the protocol: images are 224×224 and supplement state inputs, while the architecture uses a frozen DINOv2 image encoder to provide features. e04e15

What it supports. The overall change is just +0.004, from 0.438 to 0.442. RoboDesk improves by 0.125 and Meta-World by 0.069, whereas DMControl declines by 0.029 and MiniArcade by 0.004. The supported result is a domain-dependent effect under this finetuning protocol, rather than uniform benefit from images.

Where the evidence stops. The table does not present an explicitly matched state-only continuation for the extra 30M steps or uncertainty for these gains. It therefore cannot cleanly separate added visual information from additional optimization and interaction.

6.2 Results and evaluation conditions

Task & protocolReported resultComparison & interpretation
MMBench state-based multitask control

200 training tasks; default Newt; 100M total online steps plus demonstrations.

0.438; Table 9 rounds this to 0.44.

Average normalized score, higher is better

Figure 7 and Section 4.1 show higher overall performance than multitask PPO, FastTD3 and TD-MPC2. Specialist populations score 0.65 at 100M total steps and 0.80 at Table 9’s labeled 1B budget.

Shared control trails specialists. Equal online steps do not equalize demonstration-generation cost or total model capacity. e11e15e18e03

Language conditioning ablation

Full 200-task set; default-size agent; 100M total steps.

0.371 without language → 0.438 with CLIP language.

Average normalized score

Task-ID conditioning performs similarly on training tasks; removing demonstration components also reduces performance.

Task differentiation helps, without establishing compositional language understanding. e12e14

Adaptation to 20 unseen tasks and embodiments

Individual online finetuning without demonstrations; unseen instructions; five seeds in Figure 9.

Text reports 0.192 zero-shot and 0.868 after 100k steps.

Average normalized score

Prose gives scratch scores 0.013 initially and 0.480 at 100k, but Figure 9’s scratch endpoint is visibly around 0.6. Its exact final value is unresolved.

Both presentations favor finetuning; a precise gain is not consistently established. e13

Open-loop control on eight selected tasks

No feedback during 24- or 48-step plans, versus training horizon 3.

Final averages: 0.24 for four 48-step tasks; 0.60 for four 24-step tasks. Pick Screwdriver and Push Green each retain 0.83 at step 24.

Fraction of corresponding closed-loop performance

Walker Walk retains 0.25 and Assault 0.08 at step 48.

Table 8 does not support broadly matching closed-loop control in most tasks. These are performance ratios, not success rates. e17

Adding RGB observations

State-based Newt receives RGB and 30M additional finetuning steps.

0.438 → 0.442 overall (+0.004).

Average normalized score and absolute score change

RoboDesk: 0.500 (+0.125); Meta-World: 0.572 (+0.069); DMControl: 0.471 (−0.029).

Gains coexist with regressions. An explicitly matched continued-state baseline is absent, so extra training is a potential confound. e15

6.3 Ablations and diagnostic examples

Read component removals and qualitative examples within their stated evaluation conditions.

Figure 8. Capacity, task conditioning and demonstration use jointly shape the learning recipe. Original paper, p. 8 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read the four panels as separate ablation families. The bold legend entries identify the default 20M model, batch 1,024, CLIP conditioning and use of all demonstration mechanisms. The first panel compares model configurations, not merely a scalar multiplier: Table 6 changes widths, latent size, encoder depth or ensemble count between configurations. The third panel distinguishes no task language, task IDs and CLIP text embeddings. In the last panel, the division-like legend marks denote removal of demonstrations, pretraining or BC. Follow both early learning and the final plateau; similar endpoints can hide differences in how quickly useful behavior emerges. e10e12e14

What it supports. Language raises the reported aggregate score from 0.371 to 0.438, while task IDs match its training-task performance. Larger models and batches help up to the tested plateau, and the full demonstration recipe is strongest. These results support the practical combination without establishing unbounded scaling or robust interpretation of unseen language.

Where the evidence stops. The size comparison changes multiple architectural quantities, and the plot lacks uncertainty estimates. Language and task-ID parity on seen tasks does not imply equal transfer; Table 1 separately shows severe sensitivity to unseen object nouns.

7. Analysis & limitations

7.1 What the evidence leaves open

Source description

Training largely uses privileged state inputs in simulation. Atari uses raw RAM and shortened episodes, limiting transfer. One fixed instruction per training task restricts language diversity. e03e20

Source description

Table 1 exposes instruction sensitivity: unseen-object pushing succeeds at 0.3% versus 21.0% with the noun “cube”; picking reverses the trend, 10.5% versus 0%. The pushing difference is 20.7 percentage points, although the prose writes “20.7%”. e14

Reader analysis

Main comparison/ablation plots omit seed counts and uncertainty comparable to Figure 9. Table 7’s “Architecture (5M)” heading conflicts with its 20M caption and dimensions matching Table 6’s 20M column. These discrepancies remain explicit. e10e11e12e13

7.2 Questions for discussion

  1. How much of the specialist gap reflects task interference versus unequal data and total parameter capacity?
  2. Can language augmentation improve unseen-object transfer without weakening task differentiation?

8. Reproducibility audit

8.1 Requirements and known gaps

Reader analysis

Reproduction needs modified task wrappers, masks, score transforms, filtered demonstrations and frozen backbone versions. Appendix A gives domain overviews rather than every task specification; complete per-task normalization constants and a full named held-out split are not enumerated in the PDF. e02e03e04e08e13

Source description

Table 3 estimates 4.6 days for state training on two RTX 5090 GPUs, with AMD EPYC 9354 and at least 128 GB RAM; demonstrations occupy 32 GB. These are expected costs, not independently reproduced timings. Expert generation adds an estimated 1,300 GPU-hours. e16e08

Reader analysis

Proposed checks compare planning with direct-prior actions from identical checkpoints, then RGB finetuning with equal-step state-only continuation. Report absolute returns, normalization and uncertainty to separate feedback dependence and modality effects. e09e15e17e19

8.2 Proposed reproduction checks

The following checks are proposals motivated by the paper. They have not been run as part of this reading.

Check 1: Measure the contribution of search and feedback from one checkpoint

Reader-proposed experiment, not performed: use a fixed trained state-based checkpoint on the eight Table 8 tasks. From paired simulator starting states, compare direct-prior control, default closed-loop planning, and the task’s reported 24- or 48-step open-loop plan. Keep model weights and search settings fixed where applicable; record actual search work when the horizon changes. Repeat over a prespecified set of evaluation seeds and report absolute returns or success, closed-loop denominators and retained-performance ratios with uncertainty. If the advantage disappears when the checkpoint is shared, differences in learning history may explain the training-curve ablation. If repeated feedback restores performance while open-loop execution fails, the broad long-horizon claim should remain restricted. e09e17e19

Check 2: Isolate the effect of RGB from another 30M training steps

Reader-proposed experiment, not performed: branch from the same state-based checkpoint into state-only continuation and state-plus-RGB finetuning, each with 30M additional environment steps, matched task allocation, demonstrations and optimizer-update budgets. Add a control with the expanded image-input architecture but zero image features to test the effect of extra encoder parameters. Keep the specified frozen DINOv2 features and absence of frame stacking for the RGB branch. Report all domains, not only the selected rows in Table 2, with repeated-run uncertainty and compute cost. A visual-information benefit should exceed both matched controls on manipulation domains; comparable gains without image information would weaken a modality-specific explanation. e04e07e15

8.3 Reading coverage

Visual audit: Visually inspected the title, authors, affiliation and v2 date; all numbered Figures 1–26 and Tables 1–9; equations and method/training text; domain definitions, demonstration filtering, planning, architecture and hyperparameters. All six final crops were individually viewed with legends, headers and complete selected tables intact. Figure 6’s stop-gradient markers were checked against Eq. (2), and its implicit dynamics inputs against Eq. (1). The transfer plot/prose endpoint discrepancy, Table 7 architecture heading, specialist-budget caveat and open-loop claim/table tension remain disclosed. Pages 13–21 were read as text and supply no retained scientific or numerical claim in this edition. External linked resources and videos were not inspected.

PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43. Appendix coverage: reviewed.

Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.

Text reading scope & known omissions
  • Abstract and Sections 1–2: motivation and MMBench
  • Section 3: architecture, objectives and demonstration use
  • Section 4: experiments, ablations, transfer, open-loop control, visual RL and costs
  • Sections 5–6: related work and future directions
  • Acknowledgments, statements and references
  • Appendix A: all ten task domains
  • Appendix B: language instructions
  • Appendix C: demonstrations
  • Appendix D: baselines
  • Appendix E: planning
  • Appendix F: implementation and hyperparameters
  • Appendix G: open-loop control
  • Appendix H: additional comparisons
  • Appendix I: per-task results
  • Appendix J: loss curves

Outside the original text pass

  • Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
  • Identity/version: the title and three authors match the catalog. The inspected title page identifies arXiv:2511.19584v2 [cs.LG], 2 December 2025; the catalog submission date is 24 November 2025. This report covers v2; v1 was not supplied for comparison. Hao Su and Xiaolong Wang are marked as equal advisers.
  • All 11 chunks were read individually, covering all 43 PDF pages. Visual inspection covered PDF pages 1–12 and 22–43, including all numbered figures and tables; references and the appendix contents page were read as text.
  • Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout. This acquisition limitation was addressed by inspecting the supplied PDF pages and all six final crops.
  • Separate supplemental material availability has not been fully verified.
  • External code, checkpoints, demonstrations and videos were not inspected; no experiments were reproduced.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

e01PDF p. 1, title/byline/affiliation and arXiv marginInspect

Exact observed title; Nicklas Hansen, Hao Su and Xiaolong Wang; University of California San Diego; v2 dated 2 December 2025; Su and Wang share equal advising.

Go to primary source ↓
e02PDF pp. 2–4, Figure 2 and Sections 2–2.3Inspect

200 training tasks, ten domains, 20 test tasks; 159 existing plus 22 variants and 19 MiniArcade training tasks; language, state/RGB modes, masks and asynchronous wrappers.

Go to primary source ↓
e03PDF pp. 22–25, Appendix A, Table 4 and Sections A.1–A.10Inspect

Domain-specific success/return normalization, task modifications and episode lengths; 22 released MiniArcade tasks, 19 for training; Atari uses RAM, continuous-action conversion and 1,000-step episodes. Appendix A does not detail every task.

Go to primary source ↓
e04PDF p. 5, Figure 6 and Eq. (1); pp. 30–31, Appendix F architecture and normalizationInspect

Frozen CLIP and DINOv2 feed concatenative MLPs; separate encoder/dynamics/reward/value/prior components; default 512-dimensional latent and size-eight simplices; no frame stacking. Stop-gradient markers are on future observation-encoding branches.

Go to primary source ↓
e05PDF p. 5, Eq. (2) and following paragraphsInspect

Joint self-prediction/reward/TD-value learning; stopped-gradient successor encodings, log-space discrete regression, EMA target ensemble and random-subset minimum.

Go to primary source ↓
e06PDF p. 6, Eq. (3) and surrounding paragraphsInspect

Prior objective combines BC over latent rollouts, Q maximization and entropy; replay actions advance latents and supply action supervision, including planner-action distillation.

Go to primary source ↓
e07PDF p. 6, Section 3.2, items (1)–(4); p. 28, Appendix D, Model-based pretrainingInspect

All learned components pretrain with policy Q term disabled; 200,000 iterations; 50:50 demo/online sampling; retained BC and constrained planning; distributed training and compilation.

Go to primary source ↓
e08PDF p. 27, Appendix C and Table 5Inspect

4,020 demonstrations, 10–40 per task; expert training 5M steps, Atari 10M; filtering by success or >25% below median return; estimated 1,300 GPU-hours to train experts.

Go to primary source ↓
e09PDF p. 29, Appendix E and Figure 11; p. 32, Table 7, PlanningInspect

Diagonal-Gaussian CEM/MPPI-like search, elite value weighting, prior rollouts, shifted warm starts and closed/open-loop execution; constraint anneals 2M–12M; H=3, six iterations, population 512, 24 prior samples, 64 elites.

Go to primary source ↓
e10PDF pp. 31–32, Eq. (4), Tables 6–7Inspect

Default architecture; episode-length discount heuristic; 10M replay, batch 1024, update/data 0.075, Adam rates, temporal weight 0.5 and BC coefficient 10. Table 7 caption says 20M despite internal Architecture (5M) heading.

Go to primary source ↓
e11PDF p. 7, Figure 7 and Baselines; p. 8, Section 4.1, Benchmarking algorithms; pp. 27–28, Appendix DInspect

100M total-step state evaluation; Newt leads aggregate multitask comparisons but has weak domains. Baselines differ in demonstrations, conditioning or architecture. Main text labels specialists 5M steps each; main curves omit comparable seed/uncertainty details.

Go to primary source ↓
e12PDF p. 8, Figure 8 and Section 4.1 items (1)–(2); p. 9, continuation and item (3)Inspect

Ablations vary size, batch, language/task ID and demonstration uses; 0.371 without language versus 0.438 with language; task IDs match training performance; combined demonstration uses perform best.

Go to primary source ↓
e13PDF p. 8, Figure 9; p. 9, Task transferInspect

20 unseen tasks/embodiments, individual finetuning without demos, unseen instructions and five seeds. Prose gives 0.192/0.013 initially and 0.868/0.480 at 100k; scratch curve visibly ends around 0.6. Complete named held-out split is not enumerated.

Go to primary source ↓
e14PDF p. 9, Table 1 and Task transfer; p. 11, Language understandingInspect

100 trials per task on ten unseen manipulation tasks; pushing 0.3% unseen noun versus 21.0% cube, picking 10.5% versus 0%. Text identifies six pushing tasks. One fixed instruction per training task.

Go to primary source ↓
e15PDF p. 9, Table 2; p. 10, Visual observationsInspect

RGB added to state agent for 30M steps. Overall 0.438/0.442; RoboDesk 0.500 (+0.125), Meta-World 0.572 (+0.069), MiniArcade 0.437 (−0.004), DMControl 0.471 (−0.029). No explicit equal-step continued-state control.

Go to primary source ↓
e16PDF p. 9, Table 3; p. 10, Training costInspect

Expected 100M-step costs include 4.6 days on two RTX 5090 GPUs for state; AMD EPYC 9354, ≥128 GB RAM and 32 GB demonstration storage.

Go to primary source ↓
e17PDF pp. 9–10, Open-loop control and Figure 10; p. 33, Appendix G, Figure 12 and Table 8Inspect

Eight-task 24/48-step no-feedback test; final means 0.60/0.24 respectively. Pick Screwdriver and Push Green 0.83 at 24; Walker Walk 0.25 and Assault 0.08 at 48. Prose claims near-closed-loop performance in most tasks and notes drift, deceleration and stochasticity failures.

Go to primary source ↓
e18PDF p. 35, Table 9, Total row and caption; p. 27, Appendix CInspect

Newt 0.31 at 20M and 0.44 at 100M; specialist population 0.38 at 20M, 0.65 at 100M, 0.80 at labeled 1B. Table uses 5M per expert; Appendix C specifies Atari 10M. Budgets retained as labeled.

Go to primary source ↓
e19PDF p. 34, Figure 14 and captionInspect

Planning versus direct learned prior generally favors MPC, but MuJoCo favors the prior; curves motivate checkpoint-controlled evaluation.

Go to primary source ↓
e20PDF pp. 11–12, Section 6, Visual RL, Language understanding, Neural architectures and Learning strategiesInspect

Authors identify visual resource costs, limited language variation, simple MLP dynamics and heterogeneous learning rates as limitations/opportunities.

Go to primary source ↓

8.5 Primary sources

Scroll across the image to inspect details. Press Esc to close.