GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by AutoResearch

Abstract: World Action Models (WAMs) improve robot policy learning by jointly modeling actions and future visual observations, using future scene evolution as dense supervision for physically grounded action generation. However, a common design in existing WAMs is to explicitly generate future videos at inference time, incurring substantial computational overhead and hindering real-time closed-loop deployment. GigaWorld-Policy addresses this issue with an action-centered formulation, where future visual dynamics are used during training while action-only decoding is used at inference time. Building upon this framework, we present GigaWorld-Policy-0.5, an enhanced action-centered WAM designed for more efficient robot control. During pretraining, GigaWorld-Policy-0.5 adopts a mixed Action-Conditioned World Modeling (AC-WM) and WAM training strategy. This strengthens the coupling between visual dynamics and robot actions and improves the transferability of action representations for downstream policy learning. For efficient inference, GigaWorld-Policy-0.5 introduces a Mixture-of-Transformers architecture that separates visual dynamics modeling and action generation into specialized experts, reducing active computation during action-only inference and achieving 85 ms inference latency on a local RTX 4090 setup. In addition, we employ an agent-based AutoResearch pipeline to systematically search training configurations, enabling more efficient identification of optimal experimental setups while reducing the time and manual intervention required for hyperparameter tuning. Experiments and ablations show that GigaWorld-Policy-0.5 preserves the training benefits of future visual dynamics while improving inference efficiency for robot control.

• The paper introduces a Mixture-of-Transformers World Action Model that uses future visual dynamics for training while enabling lightweight action-only inference, achieving 85 ms latency on an RTX 4090 with C++ deployment.
• The paper combines Action-Conditioned World Modeling with standard WAM pretraining, improving convergence and reaching success rates of 0.85 for fruit picking and 0.89 for object placement.
• The paper reports a 0.77 average success rate on long-horizon tasks—35% higher relative to the strongest baseline—while showing that AutoResearch can automate hyperparameter selection, though its advantage over manual tuning remains unquantified.

Overview

GigaWorld-Policy-0.5 is an enhanced World Action Model (WAM) for robot manipulation that retains the training-time benefit of future visual dynamics while eliminating explicit future-video generation at inference. It builds directly on GigaWorld-Policy (Ye et al., 18 Mar 2026), whose action-centered formulation decouples future prediction between training and inference: future visual observations supervise action learning during training, but deployment uses action-only decoding. The present work contributes three main extensions. First, a Mixture-of-Transformers (MoT) architecture separates visual dynamics modeling from action generation into specialized experts, enabling a lightweight action-only inference pathway. Second, pretraining mixes Action-Conditioned World Modeling (AC-WM) with standard WAM training to strengthen the coupling between robot actions and visual state transitions. Third, an agent-based AutoResearch pipeline [karpathy2026autoresearch] automates hyperparameter search, producing the final training recipe with reduced manual intervention. The headline result is 85 ms inference latency on a local RTX 4090 with C++ deployment, alongside a real-robot success rate of 0.85 on text-following tasks and 0.77 on long-horizon tasks.

Background and motivation

WAMs jointly model robot actions and future observations, so that action representations are learned together with their visual consequences. This provides dense temporal supervision beyond the sparse action labels in standard Vision-Language-Action (VLA) training, and related work suggests such coupling yields zero-shot generalization (Ye et al., 17 Feb 2026). However, most WAM designs invoke the joint modeling process at inference, requiring explicit future-video generation, iterative denoising, or predictive rollout. Because video tokens are far more expensive than action tokens, this incurs substantial latency and risks error accumulation over long horizons. Fast-WAM (Yuan et al., 17 Mar 2026) and GigaWorld-Policy established that test-time future imagination is not required; GigaWorld-Policy-0.5 pushes this direction further on both the architecture and training-recipe axes.

Method

Formulation

The model receives multi-view observations composed of left, front, and right camera views, proprioceptive state, and a language instruction, and jointly predicts an action chunk of length $p$ and future visual observations at stride $\Delta$ with $K=\lfloor p/\Delta \rfloor$. Training uses flow matching with modality-specific flow timesteps and distinct flow-shift factors $\gamma_{\mathrm{a}}$ and $\gamma_{\mathrm{v}}$ for action and visual tokens, and the model regresses the predicted velocity field against the ground-truth flow velocity.

The central architectural invariant is the action-centered causal mask: action tokens attend to current visual tokens, state tokens, and language conditioning, but are prevented from attending to future visual tokens, while future visual tokens attend to the current context and action tokens. This prevents information leakage from future observations into action prediction, implicitly conditions visual prediction on actions during training, and makes future-visual prediction optional at deployment.

Mixture-of-Transformers architecture

GigaWorld-Policy-0.5 replaces the fully shared Transformer backbone of GigaWorld-Policy with an MoT structure comprising a visual expert and an action expert, each with its own cross-attention and FFN modules, connected through multi-modal self-attention. The visual expert uses a hidden dimension of 3072 and FFN dimension of 14336; the action expert is deliberately lightweight at 1024 and 4096 respectively. The visual expert is initialized from GigaWorld-1 (Team et al., 2 Jul 2026), a world model pretrained on over ten thousand hours of video, while the action expert is initialized from the visual-expert weights, truncating mismatched dimensions to the leading $n$ entries. Although MoT increases total parameter count, the asymmetric expert sizing reduces active computation during action-only inference.

Training pipeline

Training proceeds in two stages. The pretraining stage adapts the GigaWorld-1 world-model prior on 2K hours of filtered open-source robot data [(AgiBot-World-Contributors et al., 9 Mar 2025, Hayes et al., 15 Dec 2025, Tan et al., 17 Jul 2025), RDT-1B] plus internal real-robot data, mixing standard WAM training with AC-WM in which future visual evolution is predicted conditioned on robot actions. The post-training stage trains on target real-robot trajectories, jointly optimizing action prediction and future-visual modeling under the causal structure. Deployment uses the action-only path.

Inference acceleration

The inference stack applies KV caching (visual and language context is encoded once and reused across autoregressive action decoding), torch.compile for operator fusion and reduced dispatch overhead, and a unified C++ runtime integrating preprocessing, tensor construction, model execution, KV-cache management, and action post-processing.

Experimental results

Real-world evaluation uses an AgileX PiPER 6-DoF arm with gripper-based manipulation, comparing against $\pi_{0.5}$ (Intelligence et al., 22 Apr 2025), Motus [Motus], FastWAM (Yuan et al., 17 Mar 2026), and GigaWorld-Policy. Text-following trials use a four-stage graded score (reach, grasp, move, place, 0.25 each) over 10 trials; long-horizon tasks use binary success over 10 trials.

Task suite	$\pi_{0.5}$	Motus	FastWAM	GigaWorld-Policy	GigaWorld-Policy-0.5
Fruit picking (text following)	0.76	0.80	0.78	0.80	0.85
Object placement (text following)	0.76	0.83	0.77	0.81	0.89
Long-horizon tasks (average)	0.50	0.57	0.47	0.53	0.77

On fruit picking, GigaWorld-Policy-0.5 attains the highest score on all six instructions, with the largest margins on lemon and avocado, indicating reliable grounding of fine-grained object descriptions. On object placement, it wins on every instruction, with notable gains on compositional object-destination pairs such as fork-to-basket. The long-horizon result is the strongest claim in the paper: an average of 0.77 versus 0.57 for the best baseline, an absolute gain of 0.20 and a relative gain of 35%, with individual task scores of 0.80 on Food Heating, 0.70 on Solid Weighing, and 0.80 on Tableware Arrangement.

Inference efficiency

Method	A100 (ms)	RTX 4090 (ms)	Real-robot SR
$\pi_{0.5}$	225	110	0.76
Motus	3231	—	0.80
FastWAM	229	182	0.78
GigaWorld-Policy	360	293	0.80
GigaWorld-Policy-0.5	189	110	0.85
w/ C++ deployment	140	85	0.85

With KV cache and compilation, the model reduces A100 latency from 229 ms (FastWAM, same compiled setting) to 189 ms, a 17.5% speedup, and is faster than $\pi_{0.5}$ (225 ms). On RTX 4090 it matches $\Delta$0 at 110 ms; with the C++ runtime it reaches 85 ms, 23% faster than $\Delta$1 and 53% faster than FastWAM. The efficiency gain comes without sacrificing accuracy: the 0.85 success rate exceeds all baselines in the same comparison.

Ablations

Mixed AC-WM and WAM pretraining: compared against WAM-only pretraining under an identical post-training recipe, the mixed strategy converges faster, reaches higher success rates throughout post-training, and attains 0.85 on the fruit-picking task, with strong performance emerging at substantially earlier training steps. The implication is that explicitly modeling how actions drive visual transitions yields more transferable action representations and more sample-efficient downstream policy learning.

MoT architecture: the efficiency table above isolates the architectural contribution; the expert-separated design also simplifies initialization of the visual expert from pretrained video generation models.

AutoResearch-driven hyperparameter study: using roughly 3.9 hours (930 episodes, 300 train / 30 validation) of fruit-picking demonstrations, AutoResearch sweeps learning rates via 1K-step pilot runs, then batch sizes, then extends training steps. The learning-rate sweep selects $\Delta$2 (train action loss 0.252476, eval action MSE 0.409764), noting that $\Delta$3 achieves the lowest train visual loss (0.172330) but action quality is weighted more heavily since it directly affects execution. Batch-size alternatives do not outperform the original setting of 16. Extended training peaks in validation action MSE at 30K steps, and this checkpoint is used for real-robot evaluation. The paper does not report a controlled comparison of AutoResearch-selected hyperparameters against manually chosen ones, so the pipeline's contribution is demonstrated as a workflow rather than quantified as an accuracy gain.

Limitations and open questions

The evaluation is confined to gripper-based manipulation on a single 6-DoF arm with a small set of tabletop tasks; generalization across embodiments, dexterous hands, or mobile manipulation is not assessed. The hyperparameter study uses only about 3.9 hours of data on one task, and the AutoResearch pipeline's benefit over conventional tuning is not quantified. The AC-WM ablation is likewise restricted to a single task, so the scope of its sample-efficiency benefit remains open. The action-expert initialization by truncating visual-expert weights is a heuristic whose sensitivity to the truncation choice is not analyzed. Finally, although the causal mask prevents future-visual information from leaking into action prediction, the paper does not measure how much of the long-horizon gain is attributable to the world-model prior versus the larger pretrained initialization, leaving the relative contribution of each ingredient unresolved.

Conclusion

GigaWorld-Policy-0.5 demonstrates that the dense supervision of future visual dynamics can be retained at training time while inference cost is reduced to that of a lightweight action expert, achieving 85 ms latency on consumer-grade hardware with a 0.85 text-following success rate and a 35% relative improvement over the strongest baseline on long-horizon tasks. The mixed AC-WM pretraining and the AutoResearch pipeline are presented as complementary contributions to training efficiency and recipe reliability, though both are validated on a narrow task distribution. The paper's central empirical claim — that action-centered WAMs can match efficient VLA policies in latency while exceeding them in success rate — is supported by the reported comparisons on the evaluated manipulation suite.