RISE: Self-Improving Robot Policy with Compositional World Model
1. Paper overview
In one sentence: RISE trains a robot policy on action-conditioned imagined futures scored by a separate value model, exchanging physical exploration for compute while retaining a substantial real-data anchor. e01e02e03e04e07e09e13e18e20
| At a glance | What to know |
|---|---|
| Research problem | Source description Imitation-trained VLAs can fail after small deviations in dynamic or contact-rich tasks. Physical online RL requires serial robot interaction, supervision and resets. RISE asks whether a learned environment can supply useful new actions, states and intermediate learning signals without requiring full imagined episodes to reach terminal success. e02e03e07 |
| Core mechanism | |
| A key reported result | Dynamic Brick Sorting: 85.00%; 9.78/10 Completion success; mean stage-wise score. Physical moving-conveyor sorting; Table I; 20 autonomous trials; stage score capped at 10. RECAP: 50.00%, 9.00; π0.5: 35.00%, 8.28. The gain over RECAP is 35 percentage points. The appendix awards separate credit for grasping and correct color-bin placement; a high partial score is not full completion. e09e10 |
| Reading caution |
Core contributions
Figure 4. Separate models turn candidate actions into imagined observations and learning signals. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the upper half as preparation of two components. Task-centric batches from Agibot World and Galaxea adapt the dynamics branch; progress and temporal-difference learning adapt the value branch. Success and failure data feed task adaptation. Below the dashed divider, multiview images and a proposed action chunk enter the dynamics model. Follow the rightward arrows through predicted observations to the value model, then toward policy optimization. The text prompt is omitted in the original diagram. The reward curve is schematic: Equation (2) defines the actual chunk label by averaging predicted values and subtracting the initial value. e03e04e05e07e17e18
What it supports. The architecture separates generating a consequence from judging its usefulness. Its two learned modules can use different backbones and objectives, while the policy learns from their combined signal. This supports the dual-system classification and explains how a video model can contribute to executed robot behavior through training.
Where the evidence stops. The diagram's H-step abstraction needs implementation detail: Table VIII lists 25 predicted frames, while Table X lists 50 action steps. Their exact alignment is not specified. World-model inference here occurs inside policy training.
2. Motivation
2.1 The problem and the proposed response
Imitation-trained VLAs can fail after small deviations in dynamic or contact-rich tasks. Physical online RL requires serial robot interaction, supervision and resets. RISE asks whether a learned environment can supply useful new actions, states and intermediate learning signals without requiring full imagined episodes to reach terminal success. e02e03e07
2.2 What this reading follows
Imagine a robot choosing how to place a brick while the conveyor keeps moving. A useful training environment must predict what its particular action will do and recognize whether the outcome helps. RISE separates those jobs: a video dynamics model generates future camera views, and a value model evaluates progress and failures. Their outputs train an advantage-conditioned policy through short imagined interactions. This edition follows that information flow into the physical evaluation, then examines which component ablations support it. The supplied April 2026 revision reports strong task-specific gains, alongside unresolved differences between result tables and incomplete implementation details. e01e02e03e04e07e09e13e18e20
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 | Dual-system |
| Prediction paradigm | Other mechanisms |
| Quadrant | Outside quadrants |
3.1 Evidence-based assessment
Supports the recorded classification
The recorded WAM, Dual-system, Other mechanisms and Outside quadrants labels fit the separate policy, dynamics and value modules. Actions are proposed by the policy and evaluated through forward imagination during post-training; they are not jointly decoded with future observations or recovered by inverse dynamics. Multiview modeling is explicit, but the combined '3D multiview modeling' tag should not imply a demonstrated explicit 3D representation. e03e04e07e18
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 |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 Build a training environment from two predictions
Start with the distinction between an action and its consequence. RISE's policy proposes a chunk of robot commands from the current observation and instruction. The dynamics model receives that chunk plus visual history and synthesizes multiview observations. It does not decide whether the task is progressing. A separate value model supplies that judgment. Task-centric batching is intended to strengthen the first link: within a small group of tasks, the dynamics model sees diverse actions without every example changing scene and task simultaneously. The value branch instead starts from robot-policy features and learns progress and failure sensitivity. This division explains the compositional name: the environment's transition and evaluation functions are learned separately, then connected to produce training data for another policy. e03e04e05
Figure 5. An optimistic action request is relabeled by its predicted consequences. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the upper left: the rollout VLA receives the observation and an optimal prompted advantage. Its proposed action enters the compositional world model, which returns both a new observation and an evaluated advantage. The dashed return path reuses generated observations for another rollout; the text limits consecutive interactions to two. In the lower panel, read from the buffer on the right toward the behavior VLA and action prediction on the left. The snowflake marks the rollout role, while the flame marks the trained behavior policy; EMA transfers behavior weights back to the rollout policy. Offline examples also enter training. e03e06e07e18
What it supports. The prompted label expresses the action quality requested from the policy; the evaluated label estimates the quality actually obtained in imagination. Keeping these distinct permits learning from predicted failures as well as successes. The dynamics and value models stay frozen while policy weights change.
Where the evidence stops. The schematic calls evaluated advantage r_t and labels the loss 'MSE Loss'; Equations (2) and (8) use A and specify flow-matching training. The policy's denoising construction is not shown, so direct action-space MSE is not a complete implementation recipe.
5.2 Turn a short imagined future into an advantage
Equation (2) compares the mean value across the imagined chunk with the value of its starting observation. A favorable action therefore needs predicted progress relative to that start, rather than merely an attractive final image. This avoids waiting for an entire long task to finish in imagination. The value estimate remains the weak link if it rewards time passing while overlooking a failed grasp or stuck cover. Progress regression targets normalized episode time, whereas TD training uses successful and failed trajectories with terminal rewards of opposite sign. Figure 15 illustrates why the authors combine them. Reader interpretation: the important test is whether their combination ranks physically useful actions correctly; smoother value plots and lower video error alone cannot answer that question. e03e05e21e14
5.3 Learn from relabeled behavior, then return to the robot
During warm-up, demonstrations and corrections receive an optimal label, while recorded policy rollouts receive estimated advantage labels. This teaches a policy to condition on behavior quality. During self-improvement, the rollout policy is again asked for optimal behavior, but its imagined outcome receives a fresh evaluated label. These are not interchangeable: an optimistic request can produce a poor action. Training uses that evaluated label and the proposed action, mixed with offline experience. The behavior policy updates through flow matching, an EMA updates the rollout policy, and both world-model modules remain frozen. At deployment, only the trained action policy and asynchronous temporal ensembling are needed. The physical success table is consequently the transfer test for training in imagination, not an evaluation of online world-model planning. e06e07e10e18e09
5.4 Training and inference
During training
Dynamics pretraining uses 120k steps, batch 512 and 16 H100s for about seven days; task fine-tuning uses 50k steps, batch 64 and eight H100s for about three days. Dynamics predicts 25 frames from four context frames in three views at 192×256 resolution. e08e17
Value training uses 10k progress-only steps followed by 40k joint progress/TD steps, with discount 0.995. Both world-model modules remain frozen during policy self-improvement. Policy updates run about 10k steps with batch 64 on eight GPUs; rollout EMA decay is 0.995. e08e18
During inference
The world model is used only during learning. Deployment runs the policy with asynchronous temporal ensembling: overlapping old and new 50-step plans are linearly blended into 30 Hz, 14-dimensional commands. Figure 11 specifies two six-DoF arms, each with a one-DoF gripper. e07e10e18
5.5 Implementation flow
- Predict action consequences
Add a lightweight action encoder to GE-Base. Pretrain on Agibot World and Galaxea, then fine-tune on each task. Task-centric batches favor different actions within a small subset of tasks; stronger context noise addresses imperfect observations. The representation is generated RGB video, with no explicit 3D state specified. e04e17
- Evaluate progress and failures
Initialize a separate value model from π0.5. Progress regression supplies a dense temporal signal; TD targets add sensitivity to successful and failed outcomes. Evaluate each imagined observation and subtract the initial value from the average future value. e03e05
- Warm up advantage conditioning
Fine-tune π0.5 on task experience. Learned advantages label policy rollouts, while demonstrations and human corrections receive the optimal label. Advantages use ten uniform bins, unlike RECAP's binary conditioning. e06e18
- Improve within imagination
Start from offline states, prompt the rollout policy with optimal advantage, generate candidate actions and predicted observations, then relabel actions with evaluated advantages. Generated states feed subsequent rollouts, limited to at most two consecutive interactions. Mix these samples with offline data for flow-matching policy updates; update the rollout policy by EMA. e07e18
6. Experiments & results
RISE improves a π0.5 robot policy using short imagined interactions. A separate action-conditioned video model predicts multiview observations; a learned value model scores their progress, producing advantage labels for policy training. Real-world tests cover moving-brick sorting, backpack packing and box closing. The world model is absent from deployment, but substantial offline experience and training compute remain necessary.
6.1 Read the original evidence
Table I. Physical task completion improves across all three evaluated settings. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Each task has two columns: complete-task success and a stage-wise score capped at ten. Read a task vertically to compare policies, keeping those metrics separate. Appendix Table VII awards sorting credit for grasping and correct placement; packing and closing receive milestone scores. The appendix states that each evaluation uses twenty autonomous trials. All policy variants start from π0.5, but their adaptation data and interaction procedures differ. RECAP is reimplemented on this available backbone. For backpack packing, DAgger has higher baseline success than RECAP, so the choice of comparator changes the improvement being reported. e09e10e19
What it supports. RISE reports 85%, 85% and 95% success on sorting, packing and closing. Against RECAP's 50%, 40% and 60%, the absolute gains are 35, 45 and 35 percentage points. Backpack packing's gain over the strongest baseline success rate, DAgger's 50%, is 35 points.
Where the evidence stops. Section IV-B's stated 40% backpack margin does not match these table differences. No confidence intervals or seed variability accompany the point estimates. The results concern these task-specific physical evaluations.
Table V. Action-motion accuracy improves, while perceptual metrics expose a tradeoff. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The horizontal separator divides two experiments: fine-tuning on the authors' physical tasks and fine-tuning on Bridge. These are separate adaptation settings, not a claim of zero-shot transfer to Bridge. Follow the arrows in the header: PSNR and SSIM increase with better scores, while LPIPS, FVD and optical-flow end-point error decrease. EPE is the paper's motion-controllability proxy. Compare full RISE with the row removing task-centric batching to examine that training choice, then inspect the other metrics rather than assuming they all agree. The original bold FVD entry in the upper experiment is particularly informative. e04e14
What it supports. On the authors' tasks, RISE reduces EPE from GE's 1.05 to 0.54, and from 0.68 without task-centric batching. However, FVD changes from 61.22 without that batching to 66.84 with it. On Bridge, RISE's EPE is 0.64 versus GE's 0.96.
Where the evidence stops. The table does not support universal superiority on every metric, despite broad prose claims. Generation quality is distinct from physical control success; the supplied paper also leaves the precise evaluation split and optical-flow measurement implementation insufficiently specified.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Dynamic Brick Sorting Physical moving-conveyor sorting; Table I; 20 autonomous trials; stage score capped at 10. | 85.00%; 9.78/10 Completion success; mean stage-wise score | RECAP: 50.00%, 9.00; π0.5: 35.00%, 8.28. The gain over RECAP is 35 percentage points. The appendix awards separate credit for grasping and correct color-bin placement; a high partial score is not full completion. e09e10 |
| Backpack Packing Physical opening/insertion, lifting and zipping; Table I; 20 autonomous trials. | 85.00%; 9.50/10 Completion success; mean stage-wise score | RECAP: 40.00%, 6.13; DAgger: 50.00%, 7.00. Table arithmetic gives +45 percentage points over RECAP and +35 over the strongest baseline success rate. Section IV-B instead states a 40% margin without a matching table comparison. e09e10 |
| Box Closing Physical cup loading, flap folding and locking-tab insertion; Table I; 20 autonomous trials. | 95.00%; 9.88/10 Completion success; mean stage-wise score | RECAP: 60.00%, 8.13; π0.5: 35.00%, 7.50. RISE improves success by 35 percentage points over RECAP in this task-specific evaluation. e09e10 |
| Brick-sorting mechanism ablations Tables III–IV; physical brick sorting; the full ablation row is 70%, distinct from Table I's 85%. | No online signals: 35%; online actions only: 40%; actions and states: 70%. Complete-task success | Removing dynamics pretraining: 15%; task-centric batching: 40%; progress loss: 50%; TD loss: 35%. Generated states and both value objectives matter within this ablation setting. The paper does not reconcile its full-row performance with Table I. e12e13e09e10 |
| Dynamics prediction quality Table V, Experiment #1: fine-tuning on the authors' real-world tasks; Experiment #2 separately fine-tunes on Bridge. | RISE: 0.54 on real-world tasks; 0.64 on Bridge. Optical-flow end-point error (EPE), lower is better | GE: 1.05 and 0.96; without task-centric batching: 0.68 and 0.72. Motion error improves, but real-world-task FVD is worse with the full recipe: 66.84 versus 61.22 without task-centric batching. These are generation metrics, not robot success. e14 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table IV. Successful grasping alone does not explain reliable full-task completion. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the top pair of rows as changes to dynamics training and the next pair as changes to the value objective. The bottom row keeps all designs. The first metric checks pick-and-place success; the next checks sorting accuracy, and the final pair summarizes whole-task completion and partial score. 'w/o Progress' removes the auxiliary temporal-progress loss, while 'w/o TD Learning' removes the failure-sensitive objective. Compare each variant against this table's own full row. That row is a 70% completion result, so substituting Table I's 85% would mix reported settings that the paper does not reconcile. e05e09e10e13
What it supports. Without TD learning, pick-and-place remains 98.75%, matching the full row, yet sorting accuracy falls from 92.41% to 72.15% and completion from 70% to 35%. This supports a role for outcome-sensitive value learning beyond simply making grasps more successful.
Where the evidence stops. The ablations cover Dynamic Brick Sorting only. The full row differs from the main table, and uncertainty is not reported. These comparisons support component usefulness within the ablation setting without establishing an interaction-free causal decomposition.
Figure 15. Combining temporal progress with TD learning balances two imperfect signals. Original paper, p. 19 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the image strips first to identify the favorable cover-tucking phase and later retries at tab insertion. The three lower panels evaluate the sequence with different value-training objectives. Green shading marks the favorable interval; gray shading marks retrying behavior. Progress-only values rise smoothly and remain comparatively high during retries. The TD-only trace responds more abruptly but also swings markedly during the retry region. The combined model retains a response around the favorable interval with fewer of the TD-only trace's large swings. Use the common value-axis range to compare behavior, while remembering that none of these curves is an observed success probability. e05e13e21
What it supports. The authors use this example to motivate summing the losses: progress supplies a stable temporal scaffold, while TD supplies sensitivity to consequential events. The combined trace is still noisy. Together with Table IV, it provides qualitative and task-level evidence for the complementary objectives.
Where the evidence stops. This is a selected trajectory, not a calibration study or an aggregate variance measurement. The horizontal axis has indices but no explicit unit label. Apparent smoothness and value magnitude alone do not establish correct action rankings.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors acknowledge physically implausible transitions in rare scenarios, dependence on real data and substantial compute cost. Documented deployment failures include tracking loss, grasp slippage, zipper failures, flap misalignment and tab deformation. e20e22
Tables provide point estimates without confidence intervals or seed variability. Baselines share π0.5 initialization but differ in data and interaction protocols; RECAP is an adaptation to π0.5, not the original proprietary π0.6 system. Broad cross-task generalization is untested. e09e10e19
Table II's best tested offline ratio is 0.6 (50% completion), whereas ratios 0.1 and 0.9 give 5% and 30%. These values come from a separate sweep and cannot be merged with the 70% or 85% full-model rows. e11e13e09
7.2 Questions for discussion
- Does lower optical-flow error predict better physical advantage ranking on held-out failures?
- What explains the 50%, 70% and 85% full-recipe figures across the ratio sweep, ablations and main evaluation?
8. Reproducibility audit
8.1 Requirements and known gaps
Task data counts are 3,063 demonstrations/610 rollouts for sorting, 2,478/507 for packing, and 2,286/524 plus 540 corrections for closing. Reproduction needs these distributions and evaluation starts; success/failure counts and held-out splits are not specified. e16e10
Dynamics uses AdamW at 10⁻⁴, context noise 0.2 and 50 Euler denoising steps; value learning uses AdamW at 2.5×10⁻⁵. Clarify tasks per task-centric batch, advantage-bin boundaries, and alignment of 25 predicted frames with 50 action steps. Table X lists 'cosine' as the optimizer, leaving the actual policy optimizer unresolved. e04e17e18
Reader-proposed checks should isolate action adherence and failure-sensitive value labeling under matched data and update budgets. Compare predicted advantage against recorded outcomes before repeating policy optimization; the edition specifies two falsifiable checks. e03e13e14e21
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 better action adherence improve advantage ranking?
Reader-proposed check, not performed: train matched dynamics variants with and without task-centric batching, holding GE initialization, datasets, context noise, update counts and task fine-tuning fixed. Reserve recorded start states, action chunks and subsequent camera views as a held-out set, including incorrect-bin placements and grasp failures. With one frozen value model, compute Equation (2) on generated futures and on the corresponding recorded futures. Compare EPE, advantage-sign agreement and action ranking where comparable starts are available; keep all evaluation instances identical across variants. Lower EPE without better agreement or ranking would weaken the claim that the batching gain supplies a better policy-learning signal. e03e04e14e17e22e24
Check 2: Does TD sensitivity survive controlled policy training?
Reader-proposed check, not performed: fit progress-only, TD-only and combined value models on identical training episodes, reserving successes and subtle failures for validation. First measure how each scores failed grasps and tab-insertion retries. Then freeze a common dynamics model and train three policies from the same warm-up checkpoint with matched imagined starts, update budgets, a fixed offline mixture and a preregistered advantage-to-bin mapping. Evaluate physical completion and partial scores under the same twenty-trial protocol per seed, repeating seeds and reporting uncertainty. If the combined objective changes curves without improving held-out failure ranking or physical completion, the proposed complementary mechanism is not supported by this controlled test. e05e07e10e11e13e18e21
8.3 Reading coverage
Visual audit: Visually inspected the title/version/author page, all Figures 1–20, all Tables I–X, the method equations, evaluation protocol, limitations and appendix implementation pages. Individually inspected all six final original crops, including complete table headers, metric directions and figure labels. Figures 4–5 were cross-checked against Eqs. (1)–(2), (5)–(8) and frozen-module/EMA details; Figure 5's notation and schematic loss boundary are disclosed. References on pages 10–13 were read as text but not rendered. Separate supplements and linked code were outside the supplied material.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 14, 15, 16, 17, 18, 19, 20, 21. 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; I. Introduction
- II. Preliminary: world-model formulation and reinforcement learning
- III. Methodology: compositional world model, policy warm-up, self-improving loop, implementation
- IV. Evaluations: setup, main results, ablations
- V–VIII: related work, conclusion, limitations, acknowledgments; References
- Appendix IX–XI: additional results, evaluation/deployment, implementation and baselines
- Appendix XII–XVII: conceptual comparisons, qualitative visualizations, failure modes, VLA related work, asset licenses, broader impact
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Identity: the title page identifies arXiv:2602.11075v2 [cs.RO], 28 April 2026, with the catalog's exact title and all 13 listed authors. The catalog's 11 February 2026 submission date predates this revision. Version 1 was not supplied, so changes between versions cannot be established.
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout. This supplied extraction limitation was addressed by visually inspecting all 20 figures and all 10 tables in the retained PDF.
- Separate supplemental material availability has not been fully verified.
- Only the supplied paper was reviewed. Linked code, project pages, datasets and pretrained weights were not inspected; no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title/author/affiliation block and arXiv margin
Exact title and 13 authors match the catalog; margin identifies arXiv:2602.11075v2 [cs.RO], 28 Apr 2026. Six affiliations are printed.
Go to primary source ↓e02PDF pp. 1–2, Abstract and Section I
Motivation concerns brittle contact-rich/dynamic manipulation, physical RL costs and short-horizon imagined learning.
Go to primary source ↓e03PDF pp. 2–3, Section II-A, Eqs. (1)–(2); Section II-B, Eqs. (3)–(4)
Action-conditioned dynamics predicts multiview futures; chunk advantage averages future values minus the initial value; discretized advantage conditioning implements the policy-improvement formulation.
Go to primary source ↓e04PDF p. 4, Section III-A and Figure 4
GE-Base dynamics gains an action encoder, stronger context noise and task-centric batches over Agibot World/Galaxea; a distinct π0.5-based value model evaluates generated views.
Go to primary source ↓e05PDF pp. 4–5, Progress Value Model, Eqs. (5)–(6)
Progress loss regresses normalized episode time; TD uses zero intermediate rewards and terminal +1/−1, with summed objectives.
Go to primary source ↓e06PDF p. 5, Section III-B
Warm-up learns from demonstrations, successful/failed rollouts and corrections; learned advantages label rollouts while expert/correction data receive optimal advantage.
Go to primary source ↓e07PDF pp. 5–6, Section III-C, Figure 5 and Eqs. (7)–(8)
Optimal prompting is distinguished from evaluated advantage; offline-seeded imagination runs at most two consecutive interactions, mixes offline samples, trains by flow matching and updates rollout weights by EMA. Figure 5 schematically labels evaluated advantage r_t and loss MSE.
Go to primary source ↓e08PDF p. 6, Section III-D
Dynamics pretraining: 16 H100s, batch 512, about seven days; task tuning: eight H100s, batch 64, about three days. Value: 10k progress-only plus 40k combined steps, about one day. Policy self-improvement: about 10k steps, batch 64, eight GPUs. No world-model deployment overhead.
Go to primary source ↓e09PDF p. 6, Table I, all method/task success and score columns; Section IV-B Results
RISE success/score: sorting 85/9.78, packing 85/9.50, closing 95/9.88; RECAP 50/9.00, 40/6.13, 60/8.13. Packing DAgger is 50/7.00. The prose says a 40% packing margin.
Go to primary source ↓e10PDF p. 15, Section X-A, Table VII and Section X-B; p. 16, Figure 11
Evaluation uses 20 autonomous trials and scores capped at 10; sorting credits grasps and matched placements, packing/closing use milestones. Deployment ensembles 50-step chunks into 30 Hz, 14-dimensional control; setup has two six-DoF arms plus grippers and three cameras.
Go to primary source ↓e11PDF p. 7, Table II, Ratio and Complete columns
Offline ratios 0.1/0.3/0.6/0.9 yield completion 5/25/50/30%; ratio 0.6 has score 8.32.
Go to primary source ↓e12PDF p. 7, Table III, online-action/state indicators and Complete columns
No online action/state gives 35% completion; action only 40%; both 70%, with full score 9.43.
Go to primary source ↓e13PDF p. 7, Table IV, module variants and metric columns
Full completion is 70%; removing pretraining/task-centric batching/progress/TD gives 15/40/50/35%. Full pick-and-place is 98.75%, sorting accuracy 92.41%; without TD these are 98.75% and 72.15%.
Go to primary source ↓e14PDF p. 7, Table V, both experiments and all metric columns; p. 8, Section IV-C dynamics reliability
Real-world/Bridge EPE: RISE 0.54/0.64, GE 1.05/0.96, no-task-centric 0.68/0.72. Real-world FVD is 66.84 for RISE versus 61.22 without task-centric batching, despite broad superiority prose.
Go to primary source ↓e15PDF p. 14, Section IX, Figures 7–8 and Table VI
Advantage bins 10/5/1 yield 85/60/40% completion. Extended RECAP/DSRL training remains below RISE in the shown curves; no-pretraining dynamics has EPE 1.09 versus 0.54.
Go to primary source ↓e16PDF p. 15, Section XI-A
Task-specific data counts are sorting 3063 demonstrations and 610 rollouts; packing 2478 and 507; closing 2286 and 524 plus 540 corrections. Success/failure proportions and dataset split assignments are not given here.
Go to primary source ↓e17PDF pp. 15–16, Section XI-B; p. 17, Table VIII
Dynamics: three 192×256 RGB views, four input/25 prediction frames, pretrain/fine-tune sampling 30/15 Hz, 120k/50k steps, AdamW 10⁻⁴, 2k warm-up, context noise 0.2, and 50 Euler denoising steps.
Go to primary source ↓e18PDF p. 16, Sections XI-C–D; p. 17, Tables IX–X
Dynamics/value are frozen in self-improvement; advantages use ten uniform bins. Value uses AdamW 2.5×10⁻⁵, discount 0.995. Policy table lists batch 64, learning rate 10⁻⁴, minimum ratio 0.1, EMA 0.995, action chunk 50, dimension 14, and 'Optimizer: cosine'.
Go to primary source ↓e19PDF p. 6, Section IV-B Baselines; pp. 16–18, Section XI-E
All policies use π0.5; demonstration-only π0.5, DAgger corrections, DSRL steering and physical PPO differ in data/protocol. DSRL starts with 10 trajectories then 70 steering episodes. RECAP follows warm-up because π0.6 is inaccessible; DAgger uses single intervention.
Go to primary source ↓e20PDF pp. 8–9, Section VII
Authors identify inaccurate/underrepresented imagined dynamics, essential real-data anchoring, unresolved mixture choice and high computation as limitations.
Go to primary source ↓e21PDF p. 18, Section XIII Value Model; p. 19, Figures 14–15 and captions
Value trajectories illustrate success/failure sensitivity. Figure 15 compares progress-only smoothness, TD-only sensitivity/instability and the combined objective on box-cover tucking and tab-insertion retries.
Go to primary source ↓e22PDF p. 18, Section XIV; p. 20, Figure 18
Failure examples include incorrect sorting, tracking loss, slippage, stowing/lifting/zipper problems, incomplete cup loading, flap misalignment and tab deformation.
Go to primary source ↓e23PDF p. 18, Section XVI
Source states GE, π0.5 and Galaxea use Apache-2.0; Agibot World uses CC BY-NC-SA 4.0 and Bridge uses CC BY 4.0. These licenses were not independently checked.
Go to primary source ↓e24PDF p. 18, Section XIII Compositional World Model; p. 20, Figure 16
From a shared blue-brick state, an expert trajectory and two imagined rollouts contrast correct blue-bin placement with incorrect yellow-bin placement and corresponding reward curves.
Go to primary source ↓8.5 Primary sources
RISE: Self-Improving Robot Policy with Compositional World Model ↗
PDF · 12,697 extracted words
Source fingerprint
727d8112b18f0ff1ec796716c99b7b93597b3d4e79a157209b28550f71a97435