GigaBrain-0.5M*: a VLA That Learns From World Model-Based Reinforcement Learning
1. Paper overview
In one sentence: RAMP trains a VLA with predicted future latents and value-derived improvement labels, gaining task success while depending on corrected robot experience and incompletely specified evaluation protocols. e02e03e05e13e19
| At a glance | What to know |
|---|---|
| Research problem | Source description A policy conditioned mainly on current observations may struggle with long procedures and recovery. The authors propose using video prediction and task-progress estimates to enrich action conditioning, then improving both models with corrected real-robot experience. e02 |
| Core mechanism | |
| A key reported result | RAMP real-robot manipulation: Approximately 95 / 90 / 100, read from bar heights. Task success rate (%). Figure 15, Box Packing / Espresso Preparation / Laundry Folding; same named GigaBrain-0.5 backbone, compared after different post-training methods. RECAP: approximately 60 / 60 / 90; AWR: 75 / 55 / 90; pretrain: 60 / 40 / 85. RAMP–RECAP gaps are approximately 35, 30 and 10 percentage points, not a uniform 30% gain. Trial counts and error-bar definitions are absent. e19e02 |
| Reading caution | Reader analysis The robot comparisons omit trial counts, random seeds, detailed success criteria and the meaning of error bars. Rollout quantities, intervention budgets and baseline resource matching are insufficiently documented to establish statistical significance or isolate every cause of improvement. e18e19 |
Core contributions
- Reader analysis
The authors interpret RECAP as a marginal of RAMP over future latents. This formal relationship motivates richer conditioning; it does not itself establish better physical control. e04
Figure 1. Future-state and value predictions condition a separate action policy. Original paper, p. 1 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the lower World Model block. Its inputs include image, text and robot-state tokens, plus video and value queries. Follow the two upward routes: predicted video states enter the VLA as context, while the value route passes through N-step Estimation before becoming an advantage condition. Equation (8) specifies that transformation; the value is not itself a motor command. The upper GigaBrain-0.5 block emits robot actions. At right, the rollout arrow leads to physical experience with human intervention, and the return arrows train both systems. The detailed training-data quantities come from Section 4.1, rather than from counting the schematic’s illustrative input groups. e02e03e05e07e08e10e11
What it supports. The architecture makes the separation of responsibilities explicit: the world model forecasts scene information and task progress, while the VLA predicts executable actions. RAMP can therefore use future information at inference and still retain a deployment mode that masks those tokens and bypasses the world model.
Where the evidence stops. The diagram does not specify token packing or a gradient-freezing schedule. Standard and efficient inference are distinguished in the text; their relative end-to-end latency and success are not quantified here.
2. Motivation
2.1 The problem and the proposed response
A policy conditioned mainly on current observations may struggle with long procedures and recovery. The authors propose using video prediction and task-progress estimates to enrich action conditioning, then improving both models with corrected real-robot experience. e02
2.2 What this reading follows
A robot folding laundry needs more than a plausible next motion: it needs to recognize whether the unfolding sequence is getting closer to completion. GigaBrain-0.5M* addresses that problem with two learned systems. A video world model predicts future scene representations and values; a separate VLA converts current observations, instructions and those conditions into action chunks. RAMP then gathers physical experience with human corrections and updates the models. The figures below separate the foundation policy, the value-prediction mechanism and the subsequent RAMP gains. They also expose why the paper’s broad claims require care: some chart values disagree with the prose, and key evaluation details are missing. e02e03e05e13e19
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 separate Wan2.2 world model conditions a distinct VLA and can be bypassed at deployment, supporting Dual-system. It jointly predicts future states and values, while the policy generates actions. RAMP is advantage-conditioned post-training with latent guidance, rather than a single joint future/action predictor or explicit inverse-dynamics pipeline; Other mechanisms and Outside quadrants are appropriate. e02e03e07e10
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 Give the action policy a forecast it can use
The design keeps the action generator and the predictor separate. GigaBrain-0.5 already turns visual and language inputs into continuous action chunks, with auxiliary reasoning and trajectory outputs. RAMP adds a Wan2.2 world model whose training targets combine future visual information, value and robot state. Future observations at offsets 12, 24, 36 and 48 are encoded through a pretrained VAE; an MLP then aligns predicted future tokens with the VLA’s visual embeddings. The architectural point is that useful predictive information reaches the policy as tokens, without requiring an explicit inverse-dynamics stage. A reproduction must nevertheless resolve the representation detail: the method calls value an additional latent frame but defines channel-wise concatenation in Eq. (6). The diagram alone does not determine a tensor layout. e03e07e08
5.2 Convert progress estimates into a preference for better actions
The reward convention gives successful termination zero reward, failure a large negative penalty, and other steps minus one. Predicted values are then used in the n-step temporal-difference expression of Eq. (8). Its advantage compares accumulated rewards plus a later value against the current value. Thresholding with a strict greater-than test produces the binary improvement indicator I. Equation (3) fits both a policy without that indicator and one conditioned on it; future latents supply a richer description than the binary label alone. Deployment fixes I to one. The authors’ RECAP marginalization argument explains why they regard future conditioning as an extension. Reader interpretation: this is a motivation for the information flow, not a guarantee that imperfect forecasts improve executed actions or that the value estimates are calibrated. e04e06e08e10
5.3 Close the loop without assuming predictions are always available
RAMP collects physical trajectories containing both autonomous actions and expert corrections. The described collection software removes transition artifacts at intervention boundaries before the data are used for learning. Continual training updates the policy and also retrains the world model with rollout and base data, which the authors intend to prevent advantage collapse. During policy training, stochastic attention masking with probability 0.2 suppresses the world-model conditions; Stage 4 applies masking to both future tokens and the improvement indicator. At deployment this supports two modes: standard inference exposes a future generated with one denoising step, while efficient inference bypasses the world model and hides those tokens. The paper describes both modes but does not provide a complete paired success-versus-latency comparison, leaving their practical tradeoff unresolved. e05e08e09e10
5.4 Training and inference
During training
Detailed pretraining uses over 6,000 generated plus approximately 4,000 real-robot hours; Figure 3 labels 10,931 total. This differs from the introduction’s broader real-world-data wording. Foundation training uses batch 3,072 for 100,000 steps; task adaptation uses batch 256 for 20,000 steps. e11e12
During inference
Deployment fixes I=1. Standard mode computes future tokens using one world-model denoising step and exposes them to the action policy. Efficient mode bypasses the world model and hides future tokens. Action chunks still come from the VLA; no candidate-trajectory search or explicit inverse-dynamics decoder is specified. e03e08e10
5.5 Implementation flow
- Start with a multimodal action policy
PaliGemma-2 encodes multimodal inputs; an action DiT predicts chunks through flow matching. The foundation objective also trains autoregressive subgoal/discrete-action reasoning and GRU-decoded trajectory keypoints. Depth and 2D trajectories are optional. Knowledge Insulation is cited, without a detailed freezing schedule. e03
- Learn dynamics and progress together
Wan2.2 learns future visual latents and values from 4K real-robot hours. A pretrained VAE encodes observations at offsets 12, 24, 36 and 48. Spatially tiled values and proprioception join visual latents for flow-matching training. Prose describes additional latent frames, whereas Eq. (6) specifies channel concatenation; exact packing remains ambiguous. e06e07
- Turn predictions into policy conditions
An MLP aligns future tokens with policy visual embeddings. Values yield n-step advantages, thresholded into improvement labels. The policy fits both unconditional and improvement-conditioned action distributions; training masks world-model tokens with probability 0.2. e04e08
- Collect corrective experience and update
Physical rollouts mix autonomous execution and expert intervention. Software removes transition artifacts at intervention boundaries. Continual learning uses these trajectories for the policy and mixes rollout with base data for the world model to counter advantage collapse. Masking also covers the improvement indicator. e05e09
6. Experiments & results
GigaBrain-0.5M* adds a separate video world model to a robot VLA: predicted future latents supply scene context, while predicted values produce binary improvement labels for policy training. RAMP alternates this conditioning with physical rollouts and human corrections. Reported manipulation gains are promising, but incomplete protocols and chart–prose discrepancies limit precise interpretation.
Figure 14 contains partially clipped rightmost tick labels in its single-task Laundry Folding and multi-task Table Bussing panels. The illustrated crop selects the complete multi-task Box Packing panel at higher PDF-rendering resolution; its axes and legend remain intact. The full page was inspected, but obscured endpoint digits in the other panels cannot be verified from this supplied artifact. e18
6.1 Read the original evidence
Figure 4. The foundation policy provides a strong but uneven starting point for RAMP. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each task group using the legend: the darkest blue is GigaBrain-0.5, and the other shades represent the three foundation baselines. The vertical axis is percent success. These evaluations follow task-specific demonstration adaptation, so the chart is not a test of a single untouched zero-shot checkpoint. The two rightmost groups are particularly relevant to the later RL study: Box Packing reaches about 60%, and Espresso Preparation about 45%, leaving room for improvement. Keep this figure separate from Figure 15, whose pretrain Espresso bar is about 40%; the paper does not reconcile the evaluation difference. e12e13e19
What it supports. The plotted foundation policy is strongest or tied within each task group. Its approximate 100% Juice Preparation result ties GigaBrain-0 in the chart, while Laundry Collection is about 70%. These direct readings give a more qualified view than the prose’s claims of consistent improvements and broadly high dexterous-task success.
Where the evidence stops. Source discrepancies remain unresolved: p. 10 assigns GigaBrain-0 90% on Juice; p. 11 claims a 20% Espresso gain and groups Laundry Collection above 80%. The chart supports neither description. Trial counts are not supplied.
Figure 15. RAMP’s plotted gains over RECAP vary substantially across the three tasks. Original paper, p. 14 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the four bars in each panel as pretrain, advantage-weighted regression, RECAP and RAMP. AWR uses weighted imitation; RECAP adds an advantage condition; RAMP additionally supplies predicted future-state latents. Compare within a panel because each task has a different starting success rate. Approximate RAMP heights are 95% for Box Packing, 90% for Espresso Preparation and 100% for Laundry Folding. The corresponding RECAP bars are about 60%, 60% and 90%. Use the axis to read these as approximate percentages, and keep percentage-point differences separate from relative percentage improvements. The visible error bars have no defined statistical meaning in the accompanying text. e02e19e10
What it supports. The chart supports task-dependent improvements of approximately 35, 30 and 10 percentage points over RECAP. RAMP is the strongest plotted method on all three tasks. The large packing and espresso gains are the clearest evidence for the complete training approach; the folding comparison starts from an already high RECAP baseline.
Where the evidence stops. The abstract’s broad approximately 30% phrasing does not describe all three gaps. Trial counts, error-bar definitions, matched rollout/intervention budgets and the deployed inference mode are insufficiently specified. High plotted success does not establish failure-free operation.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| RAMP real-robot manipulation Figure 15, Box Packing / Espresso Preparation / Laundry Folding; same named GigaBrain-0.5 backbone, compared after different post-training methods. | Approximately 95 / 90 / 100, read from bar heights. Task success rate (%) | RECAP: approximately 60 / 60 / 90; AWR: 75 / 55 / 90; pretrain: 60 / 40 / 85. RAMP–RECAP gaps are approximately 35, 30 and 10 percentage points, not a uniform 30% gain. Trial counts and error-bar definitions are absent. e19e02 |
| Value prediction across eight manipulation tasks Identical pretraining data; approximately one million validation frames; task-averaged metrics. | State+value WM: 0.0621 / 0.8018 / 0.25. MAE / Kendall tau / inference seconds | VLM: 0.0683 / 0.7972 / 0.32; value-only WM: 0.0838 / 0.7288 / 0.11. Joint prediction improves reported accuracy over both controls. Value-only is fastest. These timings concern value prediction, not end-to-end robot control. e15e16 |
| Stage-2 conditioning ablation: multi-task Box Packing Four-task uniform mixture, batch 256, without rollout data; Figure 14, lower-right Box Packing panel, checkpoint visibly labelled 20,000 steps. | Approximately 50 with world-model conditioning. Task success rate (%) | Approximately 25 without conditioning. A roughly 25-point gain within trained tasks supports conditioning. Text specifies 60,000 multi-task steps while this panel ends at 20,000; some other panels have source-clipped endpoint labels. Held-out-task transfer is not demonstrated. e18 |
| Foundation policy internal evaluation Figure 4, eight tasks after task-specific adaptation. | Approximate task-order rates: 100, 90, 100, 95, 85, 70, 60, 45. Task success rate (%) | Tasks: Juice Preparation, Boxes Moving, Table Bussing, Paper Towel Preparation, Laundry Folding, Laundry Collection, Box Packing, Espresso Preparation. Chart readings conflict with some prose: Juice ties GigaBrain-0 at 100; Espresso is 45 versus pi0.5 at 35; Laundry Collection is 70. These are foundation-policy results, separate from RAMP. e12e13 |
| RoboChallenge intermediate checkpoint Reported leaderboard snapshot as of 9 February 2026; 30 standardized tasks. | 51.67 for intermediate GigaBrain-0.1. Average success rate (%) | 42.67 for pi0.5; a 9.00-percentage-point difference. The detailed section identifies the intermediate checkpoint; this is not a verified current ranking or direct evaluation of GigaBrain-0.5M*. e14 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 1. Joint state-and-value prediction improves value accuracy at an intermediate latency. Original paper, p. 13 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare the bottom row first with the middle row to test the contribution of predicting future states alongside value. Then compare it with the top VLM baseline. The three error columns have downward arrows, while Kendall’s rank correlation has an upward arrow; seconds measure inference cost. The surrounding text says the predictors share pretraining data and use approximately one million validation frames across eight tasks, with metrics averaged over tasks. It identifies an A800 GPU for the VLM’s latency. This setup evaluates prediction quality, so it should be read separately from the robot-success bars. e06e15e16
What it supports. The joint world model reports MAE 0.0621 and Kendall tau 0.8018, versus 0.0838 and 0.7288 for value-only prediction. Its 0.25-second latency lies between the value-only model’s 0.11 seconds and the VLM’s 0.32 seconds. The table supports a prediction-quality tradeoff, rather than a claim that the joint model is always fastest.
Where the evidence stops. The table does not measure complete policy latency or provide uncertainty. The conversion between the negative steps-to-completion reward formulation and the [0,1] value targets is unspecified; reproduce that convention before comparing error magnitudes.
Figure 13. A progress predictor responds to an obstruction during laundry folding. Original paper, p. 12 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow panels (1) through (4) as snapshots within one Laundry Folding sequence. The first image shows the white garment unobstructed; the second introduces a green garment, the third shows the manipulator handling it, and the fourth shows the white garment clear again. Beneath each snapshot, read the blue Progress trace against Timestep and the displayed zero-to-one scale. The orange boxes and dashed red markers focus attention on the event region. The caption interprets the drop as interference and the subsequent rise as recovery after removal. The pale continuation is not labelled as ground truth, so it should not be treated as a prediction-error reference. e06e15e17
What it supports. This example shows a useful qualitative property: the plotted value can fall when an obstruction appears and recover after the scene improves, rather than simply increasing with elapsed time. It complements the aggregate value metrics by illustrating the kind of task-progress variation the model is intended to capture.
Where the evidence stops. One selected trajectory cannot establish calibrated completion probability, reliable autonomous recovery, or the intervention rate. The [0,1] progress display is not reconciled with the negative-return value description in the method section.
Figure 14, multi-task Box Packing panel. World-model conditioning raises multi-task Box Packing success at the plotted checkpoints. Original paper, p. 13 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. This is the lower-right panel of Figure 14, showing Box Packing after training one policy on a uniform mixture of four tasks. Read the dark right-pointing triangles as using world-model conditioning and the pale diamonds as omitting it. The vertical axis measures percent success; this panel’s horizontal ticks are fully visible at 5,000, 10,000 and 20,000 training steps. Compare the two marker heights at the same checkpoint: the last pair is approximately 50% versus 25%. The surrounding text specifies Stage-2 data without generated rollout data, batch 256, and a 60,000-step multi-task schedule. It does not reconcile that schedule with this panel’s displayed endpoint. e18
What it supports. The selected panel supports a conditioning benefit on Box Packing, one of the tasks included in training, with an approximately 25-percentage-point gap at its visible 20,000-step checkpoint. On the inspected full page, the other multi-task curves also favor conditioning; the single-task Table Bussing curves tie at their final point.
Where the evidence stops. The full source figure partially clips endpoint labels in single-task Laundry Folding and multi-task Table Bussing; no hidden digits are inferred. The complete Box Packing panel preserves the 60,000-versus-20,000 schedule discrepancy. Trial counts and uncertainty are unspecified, and held-out-task transfer is untested.
7. Analysis & limitations
7.1 What the evidence leaves open
The robot comparisons omit trial counts, random seeds, detailed success criteria and the meaning of error bars. Rollout quantities, intervention budgets and baseline resource matching are insufficiently documented to establish statistical significance or isolate every cause of improvement. e18e19
Value targets are described as negative steps-to-completion, but the VLM outputs [0,1] and the diagnostic plots progress on that scale. The normalization connecting these descriptions is unspecified. A qualitative obstruction response does not establish calibrated values or reliable recovery. e06e15e17
Conditioning on an estimated future does not guarantee useful information under prediction error. The marginalization/entropy argument is a modeling rationale; the experiments do not test unseen-task generalization or fully characterize the standard-versus-efficient deployment tradeoff. e04e10e18
The authors leave more efficient model-rollout utilization and scalable autonomous data curation and updating to future work. e20
7.2 Questions for discussion
- Does future-state content improve control after holding value quality, rollout exposure and inference latency fixed?
- How much does masking preserve recovery when future predictions are wrong or unavailable?
8. Reproducibility audit
8.1 Requirements and known gaps
Required resources include the PaliGemma-2/action-DiT policy, Wan2.2 and VAE, matched robot/synthetic pretraining data, task demonstrations and HILR infrastructure. The paper specifies FSDP v2 sharding but not training GPU counts, wall time, optimizer settings, precise backbone sizes or a full module-freezing schedule. A800 is named for the VLM value-latency measurement. e03e07e09e11e12e15
A faithful rerun needs numerical reward/advantage parameters, latent packing and value normalization, action horizon/control frequency, intervention-boundary filtering, rollout/base mixing ratios and evaluation splits. Resolve the stated 60,000-step multi-task schedule versus the Box Packing panel’s visible 20,000-step endpoint before reproducing learning curves; report episode counts and uncertainty definitions. e06e07e08e09e15e18e19
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: Separate future-state information from value and training-budget effects
Reader-proposed check, not performed: repeat the four-task Stage-2 experiment with a fixed demonstration split, batch 256, identical seeds and no rollout data. Use the same pretrained value estimator and improvement labels across three policy conditions: correct future tokens, future tokens shuffled between matched task episodes, and masked future tokens. Match policy capacity and optimization steps, and report both 20,000-step and 60,000-step checkpoints with unambiguous axes. Evaluate the same randomized physical initial states, recording success counts and episode-based confidence intervals. If correct tokens consistently outperform both controls, that supports useful future-state content; if shuffled tokens match them, extra conditioning or training effects remain a competing explanation. e08e15e18
Check 2: Measure whether inference-time forecasts improve disturbance recovery
Reader-proposed check, not performed: use one final masked-training checkpoint on paired Laundry Folding trials, with and without a standardized green-garment obstruction motivated by Figure 13. Keep I=1, initial scenes and the intervention policy fixed. Compare standard one-step world-model inference with the documented bypass mode, first at a matched action-update cadence and then at each mode’s achievable cadence. Record task success, recovery time, intervention frequency and full observation-to-action latency; retain value traces around obstruction and removal. Better recovery with correct forecasts at matched cadence would support inference-time look-ahead. An advantage only at a different cadence would point to a control-rate effect; similar outcomes would leave training-time benefits as the stronger explanation. e08e09e10e17
8.3 Reading coverage
Visual audit: Visually inspected the title, authors, version, all method and experiment pages, Figures 1–15 and Table 1. Cross-checked Figure 1’s value-to-advantage route against Eq. (8), the training stages against Figure 2 and Section 3.2.2, and inference masking against pp. 7–8. Inspected all six final original crops, including the complete multi-task Box Packing panel from Figure 14 at 600 DPI. Full page 13 remains part of the inspection: its source-clipped endpoint labels in single-task Laundry Folding and multi-task Table Bussing are disclosed, and no claim reconstructs them. The 60,000-step schedule discrepancy is anchored only to the fully visible Box Packing endpoint at 20,000. Chart–prose discrepancies in foundation results and task-specific RL gains are preserved. References on pp. 15–20 were read as text; no appendix is present. External videos, code and separate supplements were outside this reading.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14. Appendix coverage: not present.
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
- Title and Abstract (p. 1)
- 1. Introduction (pp. 1–2)
- 2. Related Works, Sections 2.1–2.3 (pp. 2–3)
- 3. GigaBrain-0.5M*, Sections 3.1–3.2.2, Eqs. (1)–(8) (pp. 4–8)
- 4. Experiment, Sections 4.1–4.2 (pp. 8–14)
- 5. Conclusion and Future Work (p. 14)
- References (pp. 15–20)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Reviewed the supplied v2 artifact, marked 26 February 2026 with title-page date 27 February 2026. The catalog submission date is 12 February 2026; v1 was not supplied or compared. The observed title and author roster match the supplied catalog (e01).
- Text extraction does not reconstruct figure images; this limitation was addressed by inspecting the supplied PDF pages and original crops.
- Separate supplemental material availability has not been fully verified.
- Project videos, external benchmark documentation, code, model weights and datasets were not inspected. No experiments were reproduced.
- All six supplied text chunks were read individually. PDF pages 1–14 were visually inspected; reference pages 15–20 were read as text.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title, author block and arXiv margin
The title matches the supplied observed title, and the GigaBrain Team roster matches the catalog authors. GigaAI appears as the institutional credit. The artifact is arXiv:2602.12099v2 [cs.CV], 26 February 2026; the title-page date is 2026-2-27.
Go to primary source ↓e02PDF pp. 1–2, Abstract, Introduction and Figure 1
The authors motivate future-aware VLA control and depict a separate world model providing video states and values to a VLA, with physical human-in-the-loop rollouts feeding continual training. The abstract broadly describes approximately 30% improvement across three named tasks.
Go to primary source ↓e03PDF p. 4, Section 3.1 and Eq. (1)
GigaBrain-0.5 combines PaliGemma-2 with a flow-matching action DiT; language/discrete-action reasoning and optional GRU-decoded 2D trajectories accompany continuous action chunks. The objective combines masked token likelihood, action flow matching and trajectory regression; Knowledge Insulation is cited.
Go to primary source ↓e04PDF pp. 4–6, Section 3.2.1, Eqs. (2)–(4) and conditional-entropy argument
A KL-regularized reference-policy formulation motivates fitting unconditional and improvement-conditioned action distributions. RECAP is expressed as a marginal over future latent z. The authors argue that additional future conditioning reduces conditional action entropy.
Go to primary source ↓e05PDF p. 5, Figure 2; pp. 6–7, Section 3.2.2, Stages 1–4
RAMP consists of world-model pretraining, conditioned policy training, physical rollouts with expert intervention, and continual training. Stage 4 explicitly updates the world model with both rollout and base data as well as updating the policy.
Go to primary source ↓e06PDF p. 6, Section 3.2.2 Stage 1, Eq. (5)
Rewards are zero at successful termination, minus a large positive failure constant at failed termination, and minus one otherwise. The stated aim is prioritizing completion while reducing steps-to-completion.
Go to primary source ↓e07PDF p. 6, Section 3.2.2 Stage 1, Eqs. (6)–(7)
Future observations at offsets 12, 24, 36 and 48 are VAE encoded. Spatially tiled value and proprioception are combined with visual latents. Wan2.2 predicts latent states using flow matching and is trained with 4K hours of real robot manipulation data. Prose calls value an additional latent frame, while Eq. (6) defines channel-wise concatenation.
Go to primary source ↓e08PDF p. 7, Section 3.2.2 Stage 2 and Eq. (8)
An MLP aligns future tokens with policy visual embeddings. Predicted values enter an n-step TD advantage, thresholded into binary I with strict greater-than epsilon. Training randomly masks world-model tokens with probability 0.2; world-model inference uses a single denoising step.
Go to primary source ↓e09PDF p. 7, Section 3.2.2 Stages 3–4
Autonomous actions and expert interventions form the HILR dataset. Software removes transitional artifacts at intervention boundaries. World-model retraining with rollout and base data is intended to prevent advantage collapse; policy masking applies to both I and future tokens.
Go to primary source ↓e10PDF pp. 7–8, Inference paragraph
Deployment fixes I=1. Standard inference exposes predicted future tokens to the policy; efficient inference bypasses the world model and masks future tokens.
Go to primary source ↓e11PDF p. 8, Figure 3 and Section 4.1 Pre-training Details; p. 2, Introduction
Section 4.1 specifies over 6,000 generated and approximately 4,000 real-robot hours, with batch 3,072 and 100,000 steps. Figure 3 labels 10,931 hours, 60.85% generated and 39.15% real-world data. This differs from the introduction describing over 10K hours as real-world interactions. FSDP v2 shards all SiglipEncoderLayer modules and the first 16 Gemma2DecoderLayerWithExpert layers.
Go to primary source ↓e12PDF pp. 8–9, Section 4.1 Post-training Details; Figures 5–12, pp. 9–11
Task-specific demonstrations adapt the foundation policy to target platforms with batch 256 for 20,000 steps. Eight internal tasks and 30 RoboChallenge tasks are described. Figures show PiPER-arm and G1-humanoid deployments.
Go to primary source ↓e13PDF p. 9, Figure 4; pp. 10–11, Internal Evaluation
Approximate GigaBrain-0.5 chart heights are 100, 90, 100, 95, 85, 70, 60 and 45 percent in task order. Juice Preparation shows GigaBrain-0 at 100 despite p. 10 saying 90. Espresso shows 45 versus pi0.5 at 35 despite p. 11 claiming a 20% improvement. Laundry Collection is about 70 despite prose grouping it above 80.
Go to primary source ↓e14PDF p. 11, RoboChallenge Evaluation; p. 14, Conclusion
The detailed evaluation attributes 51.67% versus pi0.5 at 42.67% to intermediate GigaBrain-0.1, ranked first as of 9 February 2026. The conclusion broadly attributes the score to GigaBrain-0.5. The report retains the specific intermediate-checkpoint attribution.
Go to primary source ↓e15PDF pp. 11–12, Section 4.2 Value Prediction Performance
VLM and world-model value predictors use identical pretraining data and approximately one million validation frames over eight tasks. The VLM uses a learned CLS token and regression head targeting a value in [0,1], with MSE training. Metrics are averaged over tasks; p. 12 identifies A800 GPU for the 0.32-second VLM latency.
Go to primary source ↓e16PDF p. 13, Table 1, all rows and columns
VLM: latency 0.32s, MAE 0.0683, MSE 0.0106, RMSE 0.1029, Kendall 0.7972. Value-only WM: 0.11s, 0.0838, 0.0236, 0.1433, 0.7288. State-plus-value WM: 0.25s, 0.0621, 0.0099, 0.0989, 0.8018. Lower is better for the error metrics and higher for Kendall.
Go to primary source ↓e17PDF p. 12, Figure 13 and caption
Four snapshots and progress traces illustrate a green garment obstructing Laundry Folding, a predicted-value drop, removal of the obstruction, and value recovery. Orange boxes highlight the event; the plotted progress scale runs from 0 to 1.
Go to primary source ↓e18PDF p. 13, World Model Conditioning for Policy Learning and Figure 14
Four tasks use Stage-2 data without rollout data. Single-task training uses 20,000 steps, batch 256; the uniformly mixed multi-task policy is described as trained for 60,000 steps. The lower-right multi-task Box Packing panel visibly labels 5,000, 10,000 and 20,000 steps and reaches approximately 50% with conditioning versus 25% without at the last point; the prose describes an approximately 30% gap. Conditioned curves lie above unconditioned curves in all four multi-task panels. Single-task Table Bussing ties at its last point. Rightmost tick labels in the single-task Laundry Folding and multi-task Table Bussing panels are partially clipped in the supplied PDF; their hidden digits are not reconstructed.
Go to primary source ↓e19PDF pp. 13–14, Comparison with RL Baselines and Figure 15
AWR uses weighted imitation on rollouts; RECAP adds an advantage condition without state prediction; RAMP adds predicted future latents. Approximate plotted RAMP/RECAP/AWR/pretrain percentages are Box Packing 95/60/75/60, Espresso Preparation 90/60/55/40, and Laundry Folding 100/90/90/85. Error bars appear without a definition or trial counts in this comparison.
Go to primary source ↓e20PDF p. 14, Section 5 Conclusion and Future Work
Future work includes using model rollout data more efficiently, reducing computational overhead, and scaling autonomous data curation, policy refinement and world-model updating. The PDF proceeds to References on pp. 15–20 and contains no appendix.
Go to primary source ↓8.5 Primary sources
GigaBrain-0.5M*: a VLA That Learns From World Model-Based Reinforcement Learning ↗
PDF · 8,291 extracted words
Source fingerprint
4e8222c2924833d7384776d7f1a77cfc95fec318f8730235731431cbf612182a