GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by AutoResearch
1. Paper overview
In one sentence: Future-dynamics supervision and a smaller action expert support stronger robot policies with an 85 ms C++ inference path, although the study does not isolate every source of the improvement. e03e04e05e06e08e09ce10e12
| At a glance | What to know |
|---|---|
| Research problem | Source description Future-video generation can dominate WAM inference cost. The paper asks how to retain action–dynamics supervision while decoding actions fast enough for closed-loop manipulation, extending the predecessor's action-centered causal mask rather than introducing that principle anew. e02e03 |
| Core mechanism | |
| A key reported result | Long-horizon manipulation: 0.80 on each listed task; reported average 0.80. Binary success only when the full sequence completes.. Food Heating and Tableware Arrangement; 10 physical trials per task. Average 0.60 for π0.5, Motus and GigaWorld-Policy; 0.50 for FastWAM. A 20-percentage-point average gain. The prose says three tasks, but Table 3 contains only two; no third result is supplied. e08e09c |
| Reading caution |
Core contributions
Figure 2. Train with future dynamics; deploy through the action branch. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start on the left with observation, state, action and text inputs. The VAE encodes visual inputs, the state/action encoder handles robot signals, and umT5 supplies instruction conditioning. Beige visual blocks and blue action blocks connect through multimodal attention. Training includes noisy future and action tokens plus prediction losses. Move right to see the same specialized experts with a KV cache and dashed arrows for optional future processing. The decisive dependency is specified in Sections 3.1–3.2: future tokens may attend to action tokens, while action tokens cannot attend to future tokens. Thus the ordinary action output does not require the dashed future branch. e03e04e05e08
What it supports. The architecture makes future prediction a training resource that can be omitted during action decoding. Specialization also permits a smaller action expert: the implementation uses hidden dimensions 1024 for actions and 3072 for vision. These are architectural ingredients for lower active inference cost, not a latency measurement by the drawing itself.
Where the evidence stops. The small attention icon has no row/column legend, so its individual cells cannot establish the mask direction; the explanation follows the explicit text. Optional future processing must not be read as ground-truth future observations being required at deployment.
2. Motivation
2.1 The problem and the proposed response
Future-video generation can dominate WAM inference cost. The paper asks how to retain action–dynamics supervision while decoding actions fast enough for closed-loop manipulation, extending the predecessor's action-centered causal mask rather than introducing that principle anew. e02e03
2.2 What this reading follows
A robot can learn from the visual consequences of an action without rendering those consequences every time it moves. GigaWorld-Policy-0.5 develops this idea through a causal attention mask, specialized visual and action experts, and mixed robot pretraining. The key reading question is how these pieces let training use future observations while deployment relies only on present context. Its experiments combine physical manipulation, latency measurements and training diagnostics. Read their metrics carefully: fruit picking awards partial credit, long-horizon tasks require complete execution, and AutoResearch selects checkpoints using validation action error. Those measurements answer different questions about the same policy. e03e04e05e06e08e09ce10e12
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
Insufficient evidence to decide
WAM and efficient-control labels are supported. Architecture evidence shows two experts inside an integrated multimodal token sequence, so Dual-system is ambiguous if it means separately operating world-model and policy systems. Training jointly predicts actions/futures, while deployment uses direct action decoding without inverse dynamics or future rollout. Other mechanisms/Outside quadrants therefore depends on whether the taxonomy describes training or inference; the snapshot is preserved. e03e04e05e07e10
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 Why a future objective can help when the future branch is absent
Start with the attention constraint, not the deployment speed claim. During joint training, future visual tokens can use the action tokens to predict how the scene changes. The action tokens themselves cannot use the future tokens. This creates an asymmetric dependency: the future-prediction problem is conditioned on actions, but the action-prediction problem is conditioned on present observations, state and language. Reader interpretation: optimizing both losses can encourage representations useful for predicting action consequences, while the mask avoids a policy that requires unavailable future information at test time. At deployment the large future-token sequence can be omitted, leaving the action path. This is a learned action generator, not a procedure that first imagines a video and then recovers motor commands through inverse dynamics. The flow-matching objective specifies how the prediction targets are learned; it is not an online planning score. e03e04e05e06
5.2 Separate the role of expert size from the role of pretraining
The MoT architecture and mixed pretraining address different parts of the problem. The visual expert inherits GigaWorld-1 weights and has a larger hidden representation; the smaller action expert receives copied compatible weights and concentrates on action denoising. Multimodal self-attention still couples their token representations, while separate cross-attention and feed-forward modules permit specialization. Mixed pretraining then changes the learning tasks by adding action-conditioned future prediction to WAM training before target-robot adaptation. Figure 5 compares that pretraining choice under a common downstream recipe and supports faster downstream learning. Reader interpretation: it is plausible for architectural specialization to reduce inference work and for mixed supervision to improve transfer, but their evidence should not be pooled into one causal claim. Table 4 compares complete implementations, so it cannot identify the isolated speed benefit of replacing a shared backbone with MoT. e05e06e08e10e11
5.3 Model selection and robot execution answer different questions
AutoResearch screens short training runs using validation action error, retains learning rate 6×10⁻⁵ and batch size 16, and selects the 30K checkpoint during extended training. The rationale is that action prediction quality is more relevant to policy execution than visual training loss. That is a selection criterion, however, not a demonstrated equivalence between low MSE and successful manipulation. The robot tables supply the execution evidence, and even those tables use two scoring conventions: fruit and object placement award stage credit, whereas long-horizon tasks count only full completion. Reader interpretation: assess each conclusion against its own measurement. A lower validation MSE supports the selected checkpoint under that validation setup; a higher binary task score supports more completed sequences; a lower latency supports faster inference under the stated hardware/runtime conditions. None alone proves generalization to new manipulation domains. e08e09ae09be09ce10e12e13
5.4 Training and inference
During training
Initialize the visual expert from GigaWorld-1, described as pretrained on over ten thousand video hours. Copy corresponding weights into the action expert, taking leading compatible dimensions when shapes differ. Robot pretraining uses 2K hours of filtered open-source and internally collected robot data, mixing AC-WM with WAM; target-robot post-training jointly optimizes action and future-visual prediction. e05e06e08
Flow matching uses independently sampled action/visual timesteps with modality-specific shifts. Interpolated Gaussian noise and clean tokens define a velocity-regression objective; future dynamics provide auxiliary supervision. The paper does not state a frozen-module schedule or the AC-WM/WAM mixture ratio. e04e05e06
The fruit-task AutoResearch study reports approximately 3.9 hours and 930 episodes, with 300 training and 30 validation episodes. Candidates share initialization, split, optimizer and evaluation protocol. After 1K-step pilots, the selected learning rate is 6×10⁻⁵, batch size 16, and final checkpoint 30K steps. e12e13
During inference
Encode the current observation, state and instruction, reuse cached context keys/values across decoding steps, and decode actions without generating future tokens. The stack uses torch.compile and a C++ pipeline spanning preprocessing through action post-processing. Subsequent control depends on fresh robot observations; no video-to-action inverse-dynamics stage is described. e05e06e07
Section 3.4 calls decoding autoregressive, whereas Section 3.1 specifies flow-matching action chunks. Their exact scheduling relationship, denoising-step count and executed chunk prefix are not provided here. e04e07e08
5.5 Implementation flow
- Compose and encode context
Place the front camera above the left/right views in one composite image. A Wan VAE encodes vision, MLPs encode states/actions, and umT5 supplies language conditioning. e04e05
- Specialize the experts
Visual and action experts have separate cross-attention and feed-forward modules but exchange information through multimodal self-attention in a unified token sequence. Visual hidden/FFN dimensions are 3072/14336; action dimensions are 1024/4096. e05e08
- Constrain information flow
Action tokens access current context but cannot attend to future visual tokens. Future visual tokens can attend to actions. Consequently, the future objective can shape training without becoming a required input to deployed action prediction. e03e04e05
6. Experiments & results
GigaWorld-Policy-0.5 trains action prediction together with future visual dynamics, then omits future-video tokens during robot deployment. Specialized visual and action Transformer experts, mixed pretraining and runtime optimization yield stronger reported manipulation scores and an 85 ms RTX 4090 C++ inference path. The evidence is real-robot execution, but small trial counts, partial-credit metrics and incomplete implementation details limit broader conclusions.
6.1 Read the original evidence
Table 1. Fruit picking improves under a partial-credit instruction-following score. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each row as a different requested fruit under the text-following setup, then compare methods horizontally. The rightmost column reports GigaWorld-Policy-0.5; the other columns cover π0.5, Motus, FastWAM and its predecessor. Before reading the Average row as success probability, consult the protocol on pages 7–8: reaching, grasping, moving and placing each earn 0.25. The paper averages scores over 10 real-world trials in its stated evaluation protocol. A robot can therefore receive credit without completing the final placement. The table tests whether the requested object can be selected and manipulated within this setup, not whether a generated video merely looks plausible. e08e09a
What it supports. The reported average is 0.85, compared with 0.80 for Motus and GigaWorld-Policy, 0.78 for FastWAM and 0.76 for π0.5. The proposed model leads all six displayed instruction rows. This supports improved graded performance on the fruit-selection task rather than an 85% rate of fully completed trials.
Where the evidence stops. The source provides rounded aggregate scores without confidence intervals or per-trial outcomes. Do not combine this average with binary long-horizon success rates or infer robustness to unseen objects from the listed fruit categories.
Table 3. Complete-sequence performance on the two tasks actually tabulated. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Here the scoring rule changes. Pages 7–8 define long-horizon success as completing the entire sequence, with each task run 10 times on the physical robot. Read Food Heating and Tableware Arrangement as separate task rows, then compare their displayed average across methods. The proposed model reaches 0.80 on each. The baseline pattern differs by task: π0.5 has 0.50 for Food Heating and 0.70 for Tableware Arrangement, while Motus and the predecessor have 0.60 on both. The demonstrations on pages 7–8 illustrate these activities, but the quantitative claim comes from this table and its stated protocol. e08e09c
What it supports. The reported average improves from the strongest baseline average of 0.60 to 0.80, a 20-percentage-point difference. Because success is binary here, the table supports an advantage in completing these tested manipulation sequences. It does not measure the fidelity of imagined future frames or isolate a particular training component.
Where the evidence stops. The preceding paragraph says three long-horizon tasks, but the original table lists only two. No third task is reconstructed. With 10 trials per task and no uncertainty estimates, the stability of the reported gap remains unresolved.
Table 4. The 85 ms result belongs to the C++ deployment configuration. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Compare within one hardware column before moving across devices. The ordinary GigaWorld-Policy-0.5 row uses the cached, compiled setting described in the accompanying text. The following row adds C++ deployment and is the source of the 85 ms RTX 4090 headline. On that GPU, the ordinary row instead reports 110 ms, equal to π0.5. A dash for Motus means the RTX 4090 measurement is absent, not zero or a failed execution. The SR column repeats 0.85 for both proposed implementations. Read Section 3.4 alongside this crop to understand that the native pipeline includes preprocessing, model execution, cache management and action post-processing. e07e08e10
What it supports. On RTX 4090, the reported latency falls from 110 ms to 85 ms with C++ deployment; on A100 it falls from 189 ms to 140 ms. The compiled A100 row also improves on FastWAM's 229 ms. These measurements support a faster reported implementation while retaining the table's stated task score.
Where the evidence stops. Architecture, model initialization and runtime differ across method rows, so this is not an isolated MoT ablation. Timing boundaries, repeated-run variability and software versions are insufficiently specified; inference latency should not be relabeled as robot servo frequency.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| Fruit picking: text following AgileX PiPER 6-DoF arm; six fruit instructions; scores averaged over 10 real-world trials under the stated protocol. | Reported average 0.85. Graded SR: reaching, grasping, moving and placing each contribute 0.25. | π0.5 0.76; Motus 0.80; FastWAM 0.78; GigaWorld-Policy 0.80. A 0.05 absolute score gain over the strongest baselines; not an 85% complete-task success rate. e08e09a |
| Object placement: text following Same real-robot text-following protocol; bowl, fork and spoon placed on a plate or into a basket. | Reported average 0.89. Graded SR, four equally weighted stages. | Motus 0.83; GigaWorld-Policy 0.81; FastWAM 0.77; π0.5 0.76. The reported 0.06 gain over Motus concerns object–destination instructions, without a documented unseen-composition split. e08e09b |
| Long-horizon manipulation Food Heating and Tableware Arrangement; 10 physical trials per task. | 0.80 on each listed task; reported average 0.80. Binary success only when the full sequence completes. | Average 0.60 for π0.5, Motus and GigaWorld-Policy; 0.50 for FastWAM. A 20-percentage-point average gain. The prose says three tasks, but Table 3 contains only two; no third result is supplied. e08e09c |
| Action-only inference latency Table 4: A100 and local RTX 4090; compiled/cached implementation and separate C++ deployment row. | 189/110 ms on A100/RTX 4090; C++ deployment 140/85 ms, with reported SR 0.85 in both rows. Inference latency in milliseconds, lower is better. | FastWAM 229/182 ms; π0.5 225/110 ms; GigaWorld-Policy 360/293 ms. Motus has no RTX 4090 value. The 85 ms headline requires C++ deployment. Table 4 is not a controlled architecture-only ablation or a measured robot servo frequency. e07e10 |
| Mixed-pretraining ablation on fruit picking WAM-only versus mixed AC-WM/WAM pretraining, followed by the same post-training recipe. | With AC-WM: 68%, 80%, 83%, 85% at 20K/30K/40K/50K steps. Reported fruit-task SR at post-training checkpoints. | Without AC-WM: 53%, 63%, 70%, 75% at the same checkpoints. Supports faster downstream learning and a 10-point final gain; no uncertainty bands or matched pretraining-compute account are supplied. e08e11 |
| AutoResearch learning-rate selection Table 5, 1K-step pilots with the stated common split and initialization. | 6×10⁻⁵: 0.409764. Validation action MSE, lower is better. | 3×10⁻⁵: 0.416832; 4.316×10⁻⁵: 0.449387; 8×10⁻⁵: 0.461381. Best among these four pilots, not a measured advantage of agent automation over equal-budget manual search. e12e13 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 5. Mixed pretraining raises the downstream fruit-task learning curve. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow training steps from left to right, comparing the blue and orange markers at the same checkpoint. Blue denotes pretraining with AC-WM mixed into WAM; orange omits AC-WM. The source states that both variants then use the same downstream policy-training recipe on fruit picking. At 30K steps the labels are 80% and 63%; at 50K they are 85% and 75%. The y-axis begins at 50%, so the visual separation is magnified relative to a full zero-to-one scale. Although the plot labels its metric Success Rate, the fruit-task protocol elsewhere uses four-stage partial credit. e08e11
What it supports. Mixed pretraining is ahead at every plotted checkpoint, with a 17-percentage-point gap at 30K and a 10-point gap at 50K. The blue curve already reaches 80% at 30K, above the orange curve's final 75%. This is the paper's most direct evidence for improved downstream learning from AC-WM.
Where the evidence stops. The graph has no error bars, seed counts or documented matched pretraining-compute budget. It supports fewer downstream updates at the displayed scores, but cannot establish lower total training cost or a general sample-efficiency law.
Figure 6. Short pilots select a configuration; later validation selects a checkpoint. Original paper, p. 11 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Panel A is a decision trace: gray marks the baseline, green retained candidates, and red discarded candidates. Most pilots use 48 frames; the dashed 64-frame candidate is marked stopped early and should not be read as a completed low-error result. The selected configuration uses learning rate 6×10⁻⁵ and batch size 16. Panel B follows that configuration during longer training: validation action MSE decreases to 0.000113 at 30K, then rises to 0.000173 at 50K, which is discarded. The checkpoint decision therefore follows validation action error rather than the lowest visual training loss or the greatest number of updates. e12e13
What it supports. The figure explains the reported selection of a 30K-step checkpoint and shows why simply extending training need not improve the chosen validation metric. Table 5 separately identifies 6×10⁻⁵ as the best of four tested learning rates after 1K steps. Neither visual establishes that agent automation outperforms an equally resourced search strategy.
Where the evidence stops. Panel A labels the 8×10⁻⁵ pilot 0.463, while Table 5 gives 0.461381. The pilot and extended-training MSE scales also differ markedly without an explicit reconciliation. Preserve both displays and avoid computing a cross-panel improvement factor.
7. Analysis & limitations
7.1 What the evidence leaves open
Ten-trial evaluations and absent confidence intervals leave uncertainty unresolved. The paper supplies neither a broad out-of-distribution evaluation nor a dedicated failure analysis; its conclusion does not establish general-purpose manipulation. e08e09ae09be09ce14
AutoResearch leaves 600 of the stated 930 episodes unallocated. Warmup is mentioned but no schedule or warmup comparison is reported. Figure 6 labels the 8×10⁻⁵ pilot 0.463, whereas Table 5 gives 0.461381; use the table for exact pilot numbers. The very different pilot and extended-training MSE scales lack an explicit reconciliation. e12e13
7.2 Questions for discussion
- Would mixed pretraining retain its gain at equal pretraining compute and across repeated seeds?
- Does validation action MSE rank physical task completion reliably?
- How much of the latency gain comes from expert size versus runtime changes?
8. Reproducibility audit
8.1 Requirements and known gaps
Required ingredients include GigaWorld-1 initialization, Wan VAE, umT5, robot trajectories with synchronized views/state/actions/language, and the stated expert dimensions. Recover the missing mixture ratio, filtering rules, p/Δ, action normalization, flow shifts, optimizer settings, freezing policy, training hardware/compute and runtime versions before claiming full reproduction. e04e05e06e07e08e12
Proposed checks: compare mixed and WAM-only pretraining under matched data/compute and repeated seeds; separately benchmark the same action-only checkpoint with cached/compiled and C++ runtimes on an RTX 4090, controlling preprocessing and decoding settings. These would distinguish training transfer from implementation speed. e07e10e11
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 AC-WM improve transfer at a matched pretraining budget?
Reader-proposed experiment, not performed: initialize WAM-only and mixed AC-WM/WAM variants from the same weights, use identical robot trajectories and action-expert dimensions, and document the mixture ratio. Compare both matched-example and matched-GPU-time budgets, then apply one fixed downstream recipe and repeated seeds. Measure the four-stage fruit score at 20K, 30K, 40K and 50K updates, with held-out physical trials and uncertainty estimates. Keep data ordering, scoring and checkpoint selection consistent. A persistent mixed-pretraining advantage would support transferable supervision; a gain that disappears at equal compute would weaken the claim that the training objective itself explains the improvement. e05e06e08e11
Check 2: Can the optional future branch be removed without changing actions?
Reader-proposed experiment, not performed: on a fixed checkpoint and RTX 4090, hold current context, action-noise seed, precision and decoding schedule constant. Compare action decoding with optional future tokens present versus omitted, first in the same cached/compiled runtime, then repeat the action-only setting in C++. Measure action-output differences within a declared numerical tolerance and warm-run latency distributions with explicit preprocessing-to-output boundaries. Future-token changes should not alter actions if the reported mask is implemented correctly; omission should reduce computation. If actions change materially, investigate mask or cache leakage. If 85 ms is not recovered, separate runtime, precision and timing-scope differences before attributing the gap to the model. e03e04e05e07e10
8.3 Reading coverage
Visual audit: Rendered and actually inspected the title/author page, all method/training/implementation pages, evaluation definitions, all six figures, all five tables, and the conclusion. All six final original crops were viewed; the AutoResearch crop was widened and re-viewed to retain its discard annotation. Figure 2's optional arrows were checked against its caption and the explicit masking text on pages 5–6; its unlabeled mask icon was not assigned an invented cell interpretation. Pages 7–8 supply the scoring protocol underlying the result and ablation guides; pages 5–8 supply the configuration and runtime facts used by the proposed checks. Reference pages 13–17 were read in full as text, but were not visually inspected. No appendix is present and separate supplements remain unverified. This audit reports source reading, not code inspection or experimental reproduction.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12. 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 page and Abstract (p. 1)
- 1. Introduction (pp. 1–3)
- 2. Related Work, including 2.1 and 2.2 (pp. 3–4)
- 3. Method: 3.1 Overview, 3.2 Model Architecture, 3.3 Training Pipeline, 3.4 Inference Acceleration (pp. 4–7)
- 4. Experiments: metrics, implementation, 4.1 Real-World Results, 4.2 Ablation Studies (pp. 7–11)
- 5. Conclusion (p. 12)
- References [1]–[68] (pp. 13–17)
Outside the original text pass
- 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.
- All five supplied text chunks were read completely. The extraction's image omission was addressed by rendering and visually inspecting PDF pages 1–12 and all six final original crops; reference pages 13–17 were read as text only.
- Identity/version note: the exact title and identifier match. The title page displays arXiv:2607.13960v3 [cs.RO], 17 Jul 2026, and a separate report date of 2026-7-20. The catalog's submittedDate is 2026-07-17. Earlier revisions and a revision history were not supplied, so no equivalence between editions is asserted.
- Author metadata note: all 28 individually named catalog authors appear on the title page. The catalog additionally lists GigaWorld Team, which is not printed in that author list; the title block instead credits GigaAI and Tsinghua University. Metadata records the inspected individual credits and both institutions.
- No code, external project pages, prior papers or separate supplements were inspected, and no experiments were reproduced. The supplied PDF contains no appendix.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title block, alphabetical author list and arXiv margin
Exact catalog title and identifier are printed; the artifact is marked v3 with 17 Jul 2026 in the margin and 2026-7-20 above the title. It lists 28 individual authors, GigaAI and Tsinghua University, without a GigaWorld Team author label.
Go to primary source ↓e02PDF pp. 1–3, Abstract and Section 1
The authors motivate avoiding future-video cost and explicitly inherit the predecessor's action-centered training/inference separation; the new recipe adds MoT, mixed pretraining and automated search.
Go to primary source ↓e03PDF p. 4, Figure 2 and caption; pp. 5–6, Sections 3.1–3.2, causal-mask paragraphs
Training shows visual/action experts, two prediction branches and loss; inference marks future tokens/output optional and adds KV cache. Text specifies that futures attend to actions, while actions cannot attend to futures. The small mask icon has no row/column legend.
Go to primary source ↓e04PDF p. 5, Section 3.1, Eqs. (1)–(7) and intervening paragraphs
Defines composite multiview observations, action horizon p, visual stride Δ and K=floor(p/Δ); front view is above left/right. Modality-specific shifted uniform timesteps control Gaussian-noise/clean-token interpolation and velocity targets.
Go to primary source ↓e05PDF p. 6, Eq. (8) and Section 3.2
Squared velocity regression trains the model. Wan VAE, state/action MLPs and umT5 encode inputs. Experts have separate cross-attention/FFNs and multimodal self-attention in a unified sequence. Visual initialization uses GigaWorld-1; action weights copy compatible leading dimensions. Future tokens can be omitted.
Go to primary source ↓e06PDF pp. 6–7, Section 3.3
GigaWorld-1 initialization is described as using over ten thousand video hours; robot pretraining uses 2K hours of filtered open-source/internal data and adds AC-WM to WAM. Aligned target-robot trajectories support joint post-training; deployed inference generates only actions. Mixture ratio and freezing schedule are not stated.
Go to primary source ↓e07PDF p. 7, Section 3.4
Context KV caching, torch.compile and a C++ pipeline including image preprocessing, tensor construction, execution, cache management and action post-processing are described. The paragraph calls generation autoregressive without giving the flow-matching decoding schedule or software versions.
Go to primary source ↓e08PDF pp. 7–8, Section 4 Evaluation Metrics; p. 8, Implementation Details and Section 4.1 protocol
Text following assigns 0.25 to each of four stages; long-horizon SR is binary. Experiments use an AgileX PiPER 6-DoF arm and 10 trials in the stated protocols. Visual hidden/FFN dimensions are 3072/14336, action dimensions 1024/4096; chunk length and visual stride are deferred to GigaWorld-Policy.
Go to primary source ↓e09aPDF p. 8, Table 1, six fruit rows and Average row; p. 9, Section 4.1 first paragraph
Reported average graded SR is 0.85 versus π0.5 0.76, Motus 0.80, FastWAM 0.78 and predecessor 0.80. Proposed-method fruit scores are banana 0.95, apple 0.80, lemon 0.83, grape 0.93, avocado 0.78 and strawberry 0.85. No uncertainty is tabulated.
Go to primary source ↓e09bPDF p. 9, Table 2, instruction rows and Average row
Six object-placement instructions vary bowl/fork/spoon and plate/basket destinations. Reported averages are π0.5 0.76, Motus 0.83, FastWAM 0.77, predecessor 0.81 and proposed method 0.89; an unseen-composition split is not specified.
Go to primary source ↓e09cPDF p. 9, Table 3 and immediately preceding paragraph
Food Heating and Tableware Arrangement each score 0.80; their displayed average is 0.80 versus 0.60, 0.60, 0.50 and 0.60 for the four baselines. The prose says three tasks although only two task rows appear.
Go to primary source ↓e10PDF p. 10, Table 4, all latency columns; Section 4.2, Effect of MoT architecture
A100/RTX 4090 latencies: π0.5 225/110 ms, Motus 3231/unreported, FastWAM 229/182, predecessor 360/293, proposed model 189/110, proposed C++ deployment 140/85. Both proposed rows report SR 0.85. Text attributes the 189 ms setting to KV cache and compilation; no matched single-expert control is tabulated.
Go to primary source ↓e11PDF pp. 9–10, Section 4.2 mixed AC-WM/WAM ablation; p. 10, Figure 5 labels and legend
Both pretraining variants use the same downstream recipe on fruit picking. At 20K/30K/40K/50K steps, with AC-WM scores 68/80/83/85%, without scores 53/63/70/75%. The plot has no uncertainty bands; matched pretraining compute is not documented.
Go to primary source ↓e12PDF pp. 10–11, AutoResearch-driven hyperparameter study; p. 11, Table 5 and selection paragraphs
Study states 3.9 hours, 930 episodes, 300 training and 30 validation; remaining allocation is unspecified. Four 1K-step learning-rate pilots yield action MSE 0.416832, 0.449387, 0.409764 and 0.461381 for 3e-5, 4.316e-5, 6e-5 and 8e-5. The 6e-5 row has training action loss 0.252476; 3e-5 has lowest visual loss 0.172330. Final batch size is 16 and checkpoint 30K. Warmup details and a manual-search control are absent.
Go to primary source ↓e13PDF p. 11, Figure 6, Panels A/B, labels, legend and caption; Table 5
Panel A uses keep/discard markers for 1K-step pilots, labels the 8e-5 bar 0.463 rather than Table 5's 0.461381, and marks a 64-frame candidate stopped early. Panel B specifies lr=6e-5, bs=16, frames=48; action MSE labels are 0.000507/0.000113/0.000173 at 10K/30K/50K, with 30K kept and 50K discarded. The cross-panel metric-scale difference is not explained.
Go to primary source ↓e14PDF p. 12, Section 5 Conclusion; pp. 8–11, reported experiment scope
The conclusion reiterates action-only deployment and 85 ms RTX 4090 C++ latency. The reported evaluation consists of the listed manipulation tasks, latency comparisons and training studies; no dedicated limitations section, broad OOD benchmark or failure taxonomy is supplied.
Go to primary source ↓8.5 Primary sources
GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by AutoResearch ↗
PDF · 7,938 extracted words
Source fingerprint
d7b5d197ed6160a506ace836c74e7f1f467ba4f0d06432179d3f94d43ef9f297