Motus2: A Self-Evolving General World Model for Dexterous Manipulation

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

Authors: Hongzhe Bi, Zihao Zhou, Yihang Tang, Jingrui Pang, Shuhe Huang, Haitian Liu, Runqing Wang, Shuai Huang, Yichen Wang, Yiming Cheng, Ruowen Zhao, Zhenghua Li, Hengkai Tan, Xiaolong Liu, Jinhui Wan, Jiabao Liu, Min Zhao, Fan Bao, Jun Zhu

Affiliations: GensPI; Tsinghua University; BUAA; BIT

Source: 2608.30237 ↗ · Project page ↗ · Catalog record

Reading: 5 / 558 · 6 original figures & tables · ~23 min ·

1. Paper overview

In one sentence: Motus2 makes action generation, future simulation, and progress evaluation conditional interfaces of one model, enabling optional planning and policy improvement while memory and tactile feedback add distinct deployment costs. E02 E03 E04 E08 E09 E13 E15 E16 E17

At a glance	What to know
Research problem	Source description Dexterous manipulation combines expensive embodiment-specific demonstrations, visual occlusion, and contact states that vision cannot reliably reveal. The paper seeks to exploit human interaction data and unsuccessful robot experience while preventing undesirable actions from becoming imitation targets and preventing future observations from leaking into deployed action prediction. E02 E03 E18
Core mechanism	Source description A shared transformer exposes policy, action-conditioned simulator, and value interfaces through token dependencies and supervision gates. Architectural sharing is implemented through common computation, beyond merely optimizing several losses together. E02 E03
A key reported result	Five-task main manipulation suite: Midtrain-SFT: Place Ball 100%, Multi-Finger 70%, Attach Eraser 100%, Screw Bulb 90%, Put Phone 60%; average 84%. Task success rate and task-macro-average. Physical robot; 20 rollouts per task; matched target-task SFT data and evaluation. WAN-SFT, Pretrain-SFT, and Midtrain-SFT differ in initialization. Pretrain-SFT averages 51%; WAN-SFT and π0.5 each average 0%. Mid-training adds 33 percentage points over Pretrain-SFT. Reported executed-robot results support the staged initialization recipe on this suite. The 84% result belongs to supervised Midtrain-SFT, not a combined demonstration of all proposed extensions. E12 E13
Reading caution	Author claim The authors identify wearable tactile noise and morphology mismatch as the primary limitation: glove deformation can produce signals without external contact, and human gloves do not transfer directly across differently shaped robot hands. E18

Core contributions

Source description
A shared transformer exposes policy, action-conditioned simulator, and value interfaces through token dependencies and supervision gates. Architectural sharing is implemented through common computation, beyond merely optimizing several losses together. E02 E03
Author claim
The authors frame imagined branch evaluation followed by DiffusionNFT action updates as self-evolution. The implemented update freezes the video backbone and evaluator, making this a specific policy-improvement loop rather than continual updating of every interface. E06 E07
Source description
The work studies a monocular-to-stereo-to-robot curriculum, alternative visual memories, and a tactile expert that refines executable actions using recent contact measurements. E08 E09 E10 E20 E21

A shared video–action transformer and a tactile expert appear beside a policy-to-simulator-to-evaluator chain whose values feed either policy optimization or maximum-value action selection. — **Figure 1.** One shared core supports three queries, while tactile refinement forms an additional pathway. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

2. Motivation

2.1 The problem and the proposed response

Source description

Dexterous manipulation combines expensive embodiment-specific demonstrations, visual occlusion, and contact states that vision cannot reliably reveal. The paper seeks to exploit human interaction data and unsuccessful robot experience while preventing undesirable actions from becoming imitation targets and preventing future observations from leaking into deployed action prediction. E02 E03 E18

2.2 What this reading follows

Dexterous manipulation demands more than a plausible action sequence: the hands can hide the object, contact can change rapidly, and a candidate movement can make the task worse. Motus2 builds a shared video–action model that can propose actions, simulate their consequences, and score progress. Its action-first dependency structure allows ordinary control to stop after generating actions; planning activates the additional interfaces. The illustrated evidence separates the large benefit of egocentric pretraining and robot grounding from the smaller measured gains of model-based policy improvement. Separate memory and tactile studies then examine what information the controller needs when recent images are insufficient. These are different controlled studies, with different tasks and operating settings. E02 E03 E04 E08 E09 E13 E15 E16 E17

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 dimension	Recorded classification
Major category	WAMs
Architecture	One Model
Prediction paradigm	Other mechanisms
Quadrant	Outside quadrants

3.1 Evidence-based assessment

Insufficient evidence to decide

Reader analysis

WAM and One Model are supported for the core: action, video, and read-only value tokens share a transformer with explicit causal masks. WM-RL, memory, and tactile subcategories are also supported; the tactile expert has separate parameters, so sharing should not be interpreted as covering every module. Other mechanisms/Outside quadrants remain uncertain without catalog quadrant definitions: the paper combines joint video–action pretraining with action-first conditional control, simulation, and evaluation, rather than inverse-dynamics action recovery. The supplied snapshot is preserved. E02 E03 E06 E08 E09

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

Inputs	Outputs
Language instruction, current proprioception, and encoded visual observation history. Stereo egocentric video and aligned human or robot actions during training. Optional recent tactile forces and Sharpa deformation images.	Executable action chunks or tactile-refined sub-chunks. Optional action-conditioned future video latents. A categorical progress distribution and its expectation for branch ranking. Future-force predictions used as auxiliary training outputs.

4.2 Equations and their role

\begin{aligned}p_\theta(A_t,Z_t,Y_t\mid c_t)&=\pi_\theta(A_t\mid c_t)\,p_\theta^{\mathrm{wm}}(Z_t\mid c_t,A_t)\,p_\theta^{\mathrm{vm}}(Y_t\mid c_t,A_t,Z_t),\\V_\theta(c_t,A_t,Z_t)&=\mathbb E_{Y_t\sim p_\theta^{\mathrm{vm}}(\cdot\mid c_t,A_t,Z_t)}[Y_t].\end{aligned}

At is an executable action chunk, Zt its future observation latents, Yt a numerical progress-bin center, and ct the available context. The factorization specifies the action-first dependencies enforced by the mask. E02 E03

\begin{aligned}x_\sigma&=(1-\sigma)x+\sigma\epsilon,\qquad v^*=\epsilon-x,\\\mathcal L&=w_z\|v_\theta^z-v_z^*\|_2^2+\lambda_a w_a\|v_\theta^a-v_a^*\|_2^2+\lambda_v w_v\operatorname{CE}(p_\theta^{\mathrm{vm}},Y_t).\end{aligned}

For video or action targets x, independent noise levels and gates select policy, simulation, or evaluation training. The respective gate triples are (1,1,0), (1,0,0), and (0,0,1). E03

r_t=\begin{cases}\dfrac{\Delta t}{T-t},&\text{successful segments},\\-\dfrac{\Delta t}{T-t},&\text{negative segments}.\end{cases}

Δt is segment length, T trajectory length, and t segment start. This target measures signed progress relative to the remaining recorded horizon; it is not a Bellman return or calibrated completion probability. E05 E24

\begin{aligned}\hat r_i&=0.5+0.5\operatorname{clip}\!\left(\frac{V_i-\operatorname{mean}(V)}{\max(\operatorname{std}(V),\epsilon)},-1,1\right),\\v_i^+&=(1-\beta)v_{\mathrm{ref},i}+\beta v_{\theta,i},\\v_i^-&=(1+\beta)v_{\mathrm{ref},i}-\beta v_{\theta,i},\\\mathcal L_{\mathrm{NFT}}&=\mathbb E_i\!\left[\hat r_i\|v_i^+-v_i^{\mathrm{tar}}\|_2^2+(1-\hat r_i)\|v_i^--v_i^{\mathrm{tar}}\|_2^2\right].\end{aligned}

Generated branch scores determine attraction or repulsion around a fixed reference velocity. Population statistics exclude the ground-truth anchor, whose weight is one. Predicted futures affect optimization through scores and are never inputs to the action-velocity predictor. E06

5. Method in detail

5.1 Make failed experience useful without imitating it

Reader analysis

Motus2 factorizes its shared model as p(A,Z,Y | c) = π(A | c) p(Z | c,A) p(Y | c,A,Z). The context c contains instruction, proprioception, and accessible visual history; A is an action chunk, Z its future observations, and Y a categorical progress target. The action-first mask enforces this order, but supervision gates determine which trajectories teach which factor. Successful curated demonstrations activate action and video learning. Simulation mode conditions on clean recorded actions while predicting future video; evaluation mode conditions on clean actions and video while learning progress. Failed or irrelevant behavior can therefore teach transitions and unfavorable outcomes without becoming an imitation target. My interpretation is that this separation is central to the “self-evolving” claim: information about what happened and preference for what should happen enter different supervised pathways. E02 E03 E04 E05

Two attention matrices compare a joint pretraining mask with an action-first adaptation mask; context C, action A, future video Z, and value U groups have different permitted dependencies. — **Figure 2.** The decisive mask change prevents action queries from reading their own predicted future. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

5.2 Turn imagined progress into selection or a policy update

Source description

For a candidate action Aᵢ, the simulator predicts Zᵢ and the evaluator estimates a progress value Vᵢ. Best-of-N planning executes the highest-valued branch and replans after a real observation. MBRL instead uses the same values to change action-related parameters through DiffusionNFT. Scores are centered and normalized within the generated candidate group, clipped, and mapped into weights between zero and one. Higher weights emphasize a positive velocity field around an EMA reference; lower weights emphasize its negative counterpart. A ground-truth action–future pair anchors each group positively and is excluded from score normalization. The predicted future affects the action update through its score, not as an input to the action predictor. This distinction explains Table 3: candidate selection and proposal learning can help separately, but both depend on the reliability of learned simulation and evaluation. E05 E06 E07 E15 E21

5.3 Separate missing history from missing contact information

Reader analysis

A hand can obscure an object that was visible earlier, or conceal a contact force that was never observable in RGB. Motus2 addresses these as separate information problems. Global autoregression retains every past clean visual latent; hybrid memory trades some of that detail for persistent compressed tokens and a recent window. Their task-specific comparison favors global history, with cost growing over time. The tactile expert instead operates at a shorter temporal scale: the backbone denoises a 48-action chunk to an intermediate noise level once, and a small expert finishes successive six-action sub-chunks using newly acquired touch. Detached K/V features are reused across the eight refinements. Future force is predicted during training only. My reading is that these extensions support different causal explanations—remembering earlier evidence and measuring current contact—and their separate ablations should remain separate when interpreting the overall system. E08 E09 E16 E17 E22 E23

5.4 Training and inference

During training

Source description

Initialization uses Wan 2.2-TI2V-5B. Video-only monocular training runs for 500K low-resolution and 340K high-resolution steps; joint stereo video–action training follows for 450K steps. The approximately 130K corpus hours describe raw recordings before filtering, not effective training hours. E10 E20

Source description

Mid-training uses more than 100 hours of robot and supplementary alignment data. Simulation and evaluation receive 10% and 5% of samples; the remainder use policy mode. The evaluator uses 201 bins and value-loss weight 0.05. Target-robot supervised adaptation retains the action-first mask. E21

Source description

MBRL generates eight candidates per prefix plus a ground-truth anchor, with a one-chunk imagined horizon. DiffusionNFT updates only action-related parameters; the video backbone and evaluator stay frozen. A versioned EMA reference supports asynchronous rollout and optimization, with stale or excessively delayed samples discarded. E06 E07

Source description

Long-history variants train on complete episodes packed with masks that prevent cross-episode attention. Tactile losses update the expert and its modality encoders/decoder using detached backbone states. Force prediction has weight 0.1 and is auxiliary to action refinement. E08 E22 E23

During inference

Source description

Ordinary control queries only the action factor; default deployment uses five action-denoising steps. Future video and values need not be generated. After execution, actual observations refresh the policy context. E04 E08 E21

Source description

Planning proposes N chunks, simulates and scores each, executes the highest-valued chunk, and replans after receiving the real observation. Tactile deployment outputs refined actions; it does not require future-force prediction. E06 E09

5.5 Implementation flow

Standardize human interaction data
Synchronize recordings, wrist poses, and language annotations. A 134-dimensional representation combines bilateral wrist poses with wrist-local finger keypoints. Human fingers are retargeted offline to first-generation 20-DoF Wuji joints using constrained geometric optimization; target-robot post-training instead uses recorded native joint angles. E10 E11
Construct the shared causal computation
Language enters through cross-attention; stereo video, proprioception, actions, and value queries participate in the shared transformer. Pretraining allows mutual video–action visibility within a chunk. Robot adaptation blocks action tokens from reading that chunk’s future video or value query. Video can read actions, and the read-only value query can read both. E02 E03
Route supervision by trajectory quality
Curated successful trajectories supply action targets. Simulation mode conditions on clean recorded actions and learns future video; evaluation mode conditions on clean actions and video and learns progress. Failed, suboptimal, and task-irrelevant interactions enter applicable simulation or evaluation modes without supervising imitation. E03 E04
Learn a branch evaluator
Successful segments receive positive relative-progress targets and negative trajectories receive their signed counterpart. A categorical readout predicts discretized targets, and its expectation ranks candidate action–future pairs. These values are explicitly not calibrated success probabilities. E02 E05 E24
Retain observation history
The default sliding window stores actual observed latents, evicts old entries, and rebases temporal RoPE. Global autoregression retains every preceding observation at growing cache and attention cost. Hybrid memory retains initial anchors and recent frames at full resolution while compressing intermediate history into persistent tokens. E08 E22
Refine actions with tactile feedback
The backbone partially denoises a 48-action chunk to σc = 0.2 and supplies detached layer-wise caches. At 30 Hz action execution, the tactile expert refines eight successive six-action sub-chunks using the latest tactile window. Future-force tokens can read actions, but action refinement cannot read noised future forces. E09 E23

6. Experiments & results

Motus2 adapts a jointly pretrained video–action transformer into an action-first controller: propose an action chunk, optionally simulate its visual consequences, and evaluate the branch using learned relative progress. Best-of-N planning selects branches, while DiffusionNFT updates action parameters using the same scores. Separate studies examine full-history context and tactile refinement. The main physical-task suite reaches 84% average success after robot-domain mid-training; policy improvement and the other extensions are evaluated separately. [E02, E06, E13, E15–E17]

6.1 Read the original evidence

Five real-robot task columns compare π0.5, WAN-SFT, Pretrain-SFT, and Motus2 Midtrain-SFT; average success is zero, zero, 51, and 84 percent respectively. — **Table 2.** Egocentric pretraining and robot-domain grounding are separate steps in the main comparison. Original paper, p. 12 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

6.2 Results and evaluation conditions

Task & protocol	Reported result	Comparison & interpretation
Five-task main manipulation suite Physical robot; 20 rollouts per task; matched target-task SFT data and evaluation. WAN-SFT, Pretrain-SFT, and Midtrain-SFT differ in initialization.	Midtrain-SFT: Place Ball 100%, Multi-Finger 70%, Attach Eraser 100%, Screw Bulb 90%, Put Phone 60%; average 84%. Task success rate and task-macro-average	Pretrain-SFT averages 51%; WAN-SFT and π0.5 each average 0%. Mid-training adds 33 percentage points over Pretrain-SFT. Reported executed-robot results support the staged initialization recipe on this suite. The 84% result belongs to supervised Midtrain-SFT, not a combined demonstration of all proposed extensions. E12 E13
Policy optimization and test-time selection Put Phone and Multi-Finger; physical robot; 20 rollouts per task; factorial comparison from the same target-robot checkpoint.	Baseline: 60%/70%, average 65.0%; planning: 65%/70%, average 67.5%; MBRL: 65%/80%, average 72.5%; both: 70%/80%, average 75.0%. Success rate	Planning adds 2.5 percentage points with either policy; MBRL adds 7.5 points with either inference setting. The measured improvements distinguish changing candidate selection from changing policy weights. Their statistical reliability is unresolved because uncertainty estimates are not reported. E12 E15
Long-horizon context probes in simulation Find Square and Press Button; 25 rollouts per task; matched training and paired configurations.	Global autoregression: 84%/72%, average 78%. Success rate	Hybrid memory: 64%/40%, average 52%. Full observation retention outperforms the tested compressed representation in simulation. This table does not include the default sliding-window baseline. E16 E22
Long-horizon context probes on the physical robot Find Square and Press Button; 20 rollouts per task; matched training and paired configurations.	Global autoregression: 65%/50%, average 57.5%. Success rate	Hybrid memory: 30%/20%, average 25.0%. The ordering agrees with simulation, but these physical execution rates must remain separate from the simulation results. E12 E16
Contact-sensitive manipulation Pull Out Paper Cup with Sharpa Wave and Tear Paper with Wuji Hand 2; 20 physical rollouts per task; matched checkpoint and post-training data.	With tactile expert: 75%/70%, average 72.5%. Success rate	Without tactile expert: 65%/55%, average 60.0%; average gain 12.5 percentage points. This establishes a benefit for the complete tactile package. It does not isolate the contributions of refinement, auxiliary force prediction, or added model capacity. E17 E23
Stereo human-action prediction scaling Nested 2k, 4k, 10k, and 20k raw-hour subsets; common trajectory-disjoint held-out set; best validation epoch selected at each scale.	*Reported fitted relation: Lval(D) ≈ 0.101 − 0.005 ln D.** MSE between the mean sampled action prediction and normalized ground-truth wrist/hand actions	The text reports monotonically decreasing best validation error across the four scales. This is a fitted human-action prediction trend over the measured range, not an executed-robot success result. Individual curve values are omitted because the figure was not visually inspected. E14

6.3 Ablations and diagnostic examples

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

A four-row factorial comparison toggles planning and model-based reinforcement learning on Put Phone and Multi-Finger, producing average success rates of 65, 67.5, 72.5, and 75 percent. — **Table 3.** Planning changes candidate selection; model-based learning changes the proposal policy. Original paper, p. 14 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

Simulation and real-robot blocks compare hybrid working memory with global autoregression on Find Square and Press Button; full history has higher success in all four task-setting comparisons. — **Table 4.** Keeping all observations wins these memory probes, with a growing history cost. Original paper, p. 14 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

Tactile feedback ablation reports paper-cup success increasing from 65 to 75 percent and tear-paper success increasing from 55 to 70 percent, with a macro-average increase from 60 to 72.5 percent. — **Table 5.** Fresh tactile refinement helps on two contact-sensitive tasks. Original paper, p. 15 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

7. Analysis & limitations

7.1 What the evidence leaves open

Author claim

The authors identify wearable tactile noise and morphology mismatch as the primary limitation: glove deformation can produce signals without external contact, and human gloves do not transfer directly across differently shaped robot hands. E18

Reader analysis

The physical studies use 20 rollouts per task without reported confidence intervals or repeated-training-seed estimates. Separate task suites test MBRL, memory, and tactile sensing, so the reported gains do not establish the performance of a single configuration combining every extension. E12 E13 E15 E16 E17

Reader analysis

Self-evolution is demonstrated through policy updates with frozen dynamics and value components. The experiments do not quantify simulator error, imagined-branch ranking accuracy, or sustained improvement across repeated real-world data-collection cycles. Shared model biases could therefore influence both imagined outcomes and their scores. E07 E15 E24

Reader analysis

Signing targets according to trajectory outcome can label useful prefixes of failed executions negatively. Whether this penalizes recoverable behavior is untested. The qualitative evaluator examples use executed actions and observed video, which does not directly establish reliability on generated futures. E05 E24

Reader analysis

The memory comparison lacks a sliding-window baseline and measured latency or memory costs. The tactile comparison changes both refinement and prediction machinery, leaving the auxiliary prediction objective's independent effect unresolved. E08 E16 E17 E22

Reader analysis

Qualitative evidence has internal documentation mismatches. Section 5.4 describes Screw Bulb examples that differ from Appendix F's four-task account. Appendix F also describes phone-box closure and separate-container sorting, whereas Appendix G specifies a phone stand and an orange basket. Those trajectories cannot be assumed to validate the main-suite task checkers. E25

7.2 Questions for discussion

Would segment-level recovery labels improve over signing every segment by trajectory outcome? [E05, E24]
How accurately do evaluator rankings on imagined branches predict real execution outcomes before and after NFT optimization? [E06, E15]
Can hybrid memory approach full-history success under matched latency, and does force prediction help beyond tactile refinement alone? [E16, E17, E22]

8. Reproducibility audit

8.1 Requirements and known gaps

Source description

The supplied title, complete author list, and identifier match the catalog; the cover identifies arXiv:2608.30237v1, dated 31 August 2026. This review uses the supplied artifact provenance. No code was inspected and no experiments were reproduced. E01

Reader analysis

Reproducing the full curriculum requires the Wan initialization, access to the mixed open and procured corpus, synchronized pose processing and retargeting, robot/alignment trajectories, and the relevant bimanual hardware. Source-specific filtering thresholds, exact target-task data quantities, and complete training compute are not specified sufficiently for an exact rerun. E10 E11 E19 E20 E21

Source description

Several settings are concrete: default action sampling uses five steps, while the memory comparison uses 50; MBRL rollout action/video sampling uses four/eight steps, β = 0.1, EMA decay 0.99, publication every ten updates, and a 600-second sample-age cutoff. Comparisons must preserve these distinct protocols rather than silently sharing one sampler configuration. E07 E21 E22

Reader analysis

A focused reproduction would repeat the policy-optimization/planning factorial study from one checkpoint with paired physical configurations, then test whether branch rankings agree with actual outcomes. This directly probes the proposed feedback mechanism without requiring a new large-scale pretraining run. E06 E15

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: Does predicted progress rank real consequences?

Proposed check: freeze the policy and generate the same candidate set for every matched task prefix. Compare learned-value ranking with random ranking, shuffled values, and an offline oracle based on realized outcomes in resettable evaluation states. Separately score real future observations and model-predicted futures to localize simulator versus evaluator error. Report pairwise ranking accuracy, calibration, success, and compute at a fixed candidate count. A learned ranking advantage that survives matched sampling would support the planning mechanism; systematic high scores for failed imagined branches would reveal exploitable model error. E05 E06 E15 E24

Check 2: What history benefit survives a fixed inference budget?

Proposed check: compare the bounded sliding-window default, hybrid memory, and global autoregression using identical demonstrations, denoising steps, and paired Find Square/Press Button resets. Vary the delay between informative observation and required action, recording peak KV memory and control latency alongside success. Repeat with a strict per-decision compute budget. The falsifiable prediction is that full history helps increasingly after evidence leaves the recent window, but may lose its advantage when attention growth reduces feedback frequency. This would distinguish information retention from the operating budget used in the reported table. E08 E16 E21 E22

8.3 Reading coverage

Visual audit: Original PDF pages 4, 6, 12, 14, and 15 and all six final crops were visually inspected. Figure 1 interfaces, Figure 2 attention cells, Table 2 main results, Table 3 factorial planning/MBRL comparison, Table 4 memory study, and Table 5 tactile study were read with their source protocols. Every table header and denominator context was checked; cropped graphics retain their original content. The PDF SHA-256 was recomputed and matches the base report. Other figures, external demonstrations, code execution, value calibration, and robot replication remain outside this illustrated pass.

PDF pages inspected for this edition: 4, 6, 12, 14, 15. 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
1. Introduction
2. Related Work
3. Method
3.1. Overview
3.2. General World Modeling
3.3. Value-Guided Closed-Loop Self-Evolution
3.3.1. Progress-Based Value Learning
3.3.2. Planning and Policy Optimization
3.4. Working Memory
3.5. Tactile Expert: Tactile Refinement and Prediction
4. Large-Scale Egocentric Dataset
5. Experiments
5.1. Experiment Setup
5.2. Main Results
5.3. Scaling Laws of Stereo Egocentric Human Data
5.4. Model-Based Policy Improvement and Test-Time Planning
5.5. Long-Horizon Context Mechanisms
5.6. Effect of Tactile Feedback
6. Conclusion
7. Limitations and Future Work
Acknowledgments
References
A. Robot System and Teleoperation Stack
A.1. Robot Platforms
A.2. Teleoperation
B. Human–Robot Alignment Data Collection
C. Egocentric Data Curation and Offline Annotation
C.1. Pose Standardization
C.2. Quality-Aware Temporal Processing
C.3. Language Annotation
D. Hand Retargeting Algorithm
E. Model and Training Details
E.1. Pre-training
E.2. Mid-training
E.3. Post-training
F. Qualitative Value-Model Trajectories
G. Evaluation Tasks

Outside the original text pass

Review covers the supplied text, including appendix prose and figure captions; no PDF figure images were visually inspected.
Numerical curves were not reconstructed from extracted figure labels.
Separate supplemental material availability and completeness remain unverified.
Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
Separate supplemental material availability has not been fully verified.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

E01PDF p. 1, title, author block, and arXiv version lineInspect

The title and ordered author list match the catalog. The cover identifies arXiv:2608.30237v1, 31 August 2026, and lists a project-page link.

1. Paper overview

Core contributions

2. Motivation

2.1 The problem and the proposed response

2.2 What this reading follows

3. Research context

3.1 Evidence-based assessment

4. Problem formulation

4.1 Inputs and outputs

4.2 Equations and their role

5. Method in detail

5.1 Make failed experience useful without imitating it

5.2 Turn imagined progress into selection or a policy update

5.3 Separate missing history from missing contact information

5.4 Training and inference

During training

During inference

5.5 Implementation flow

6. Experiments & results

6.1 Read the original evidence

6.2 Results and evaluation conditions

6.3 Ablations and diagnostic examples

7. Analysis & limitations

7.1 What the evidence leaves open

7.2 Questions for discussion

8. Reproducibility audit

8.1 Requirements and known gaps

8.2 Proposed reproduction checks

Check 1: Does predicted progress rank real consequences?

Check 2: What history benefit survives a fixed inference budget?

8.3 Reading coverage

Outside the original text pass

8.4 Traceable evidence

8.5 Primary sources

Motus2: A Self-Evolving General World Model for Dexterous Manipulation ↗