GigaWorld-Policy: An Efficient Action-Centered World--Action Model

Abstract: World-Action Models (WAM) initialized from pre-trained video generation backbones have demonstrated remarkable potential for robot policy learning. However, existing approaches face two critical bottlenecks that hinder performance and deployment. First, jointly reasoning over future visual dynamics and corresponding actions incurs substantial inference overhead. Second, joint modeling often entangles visual and motion representations, making motion prediction accuracy heavily dependent on the quality of future video forecasts. To address these issues, we introduce GigaWorld-Policy, an action-centered WAM that learns 2D pixel-action dynamics while enabling efficient action decoding, with optional video generation. Specifically, we formulate policy training into two coupled components: the model predicts future action sequences conditioned on the current observation, and simultaneously generates future videos conditioned on the predicted actions and the same observation. The policy is supervised by both action prediction and video generation, providing richer learning signals and encouraging physically plausible actions through visual-dynamics constraints. With a causal design that prevents future-video tokens from influencing action tokens, explicit future-video generation is optional at inference time, allowing faster action prediction during deployment. To support this paradigm, we curate a diverse, large-scale robot dataset to pre-train an action-centered video generation model, which is then adapted as the backbone for robot policy learning. Experimental results on real-world robotic platforms show that GigaWorld-Policy runs 9x faster than the leading WAM baseline, Motus, while improving task success rates by 7%. Moreover, compared with pi-0.5, GigaWorld-Policy improves performance by 95% on RoboTwin 2.0.

• The paper introduces a shared 5B-parameter diffusion Transformer that uses future visual dynamics as training supervision while decoding actions without mandatory video rollouts, reducing inference time to 360 ms from Motus’s 3,231 ms.
• GigaWorld-Policy achieves 0.85 success under randomized RoboTwin 2.0 conditions and 0.83 average real-world success across four PiPER-arm tasks, while improving over major VLA baselines at comparable speeds.
• The paper’s curriculum combining web-video initialization with roughly 10,000 hours of embodied pre-training raises success from 0.45 when trained from scratch to 0.83 and matches π0.5’s peak performance with 10% of its demonstrations.

GigaWorld-Policy is a World–Action Model (WAM) that restructures how video-generation backbones are adapted for robot policy learning. Rather than requiring explicit future-video rollouts at inference time, it treats visual-dynamics prediction as a training-time supervision signal and an optional inference-time branch, enabling low-latency action decoding from a single shared diffusion Transformer. The system is built on a 5B-parameter Wan 2.2 backbone, pre-trained on roughly 10,000 hours of curated embodied video, and evaluated on RoboTwin 2.0 simulation and real-world manipulation tasks on AgileX PiPER arms.

Motivation: supervision sparsity and inference overhead

The paper identifies two bottlenecks in existing approaches. First, VLA models trained purely by imitation learning suffer from sparse, low-diversity action supervision relative to high-dimensional observations, which encourages contextual shortcuts and collapse into a small set of action prototypes rather than physically consistent action distributions. Second, prior WAMs such as Motus (Bi et al., 15 Dec 2025), Cosmos-Policy (Kim et al., 22 Jan 2026), VideoVLA (Shen et al., 7 Dec 2025), and two-stage pipelines like Mimic-video (Pai et al., 17 Dec 2025) jointly model future visual dynamics and actions but require iterative diffusion sampling to roll out future videos at inference, incurring high latency (Motus requires 3,231 ms per step on an A100). Errors in video prediction can also propagate into action decoding and compound over long horizons.

Architecture and causal masking

GigaWorld-Policy packs all modalities into one token sequence $T_t = [T_o; T_s; T_a; T_f]$ — current multi-view observation tokens, proprioceptive state tokens, action-chunk tokens, and future-observation tokens — processed by a single shared stack of Transformer blocks with shared QKV projections, preserving the computational profile of the pre-trained backbone. Multi-view inputs (left/front/right cameras) are merged into a single composite image so cross-view reasoning occurs without modifying the backbone. Future frames are predicted sparsely at stride $\Delta$ along the action horizon to reduce redundancy.

A blockwise causal attention mask enforces the key factorization: state and observation tokens attend only to each other; action tokens attend to $\{T_s, T_o\}$ only; future-video tokens attend to $\{T_s, T_o, T_a\}$, implementing action-conditioned feedforward dynamics. Language conditioning is injected via cross-attention outside the causal ordering. This design has two consequences: it prevents information leakage from predicted future frames into action generation, and it makes the video branch structurally optional at inference — the model can decode actions alone without instantiating any future-video tokens, or optionally denoise video tokens using a KV cache saved during action decoding.

Training uses flow matching on both modalities, with a combined objective $$\mathcal{L}_{all} = \lambda_{video}\mathcal{L}_{video} + \lambda_{action}\mathcal{L}_{action}$$ weighted $\lambda_{action}=5$, $\lambda_{video}=1$ during post-training. Action chunks have length $p=48$ with stride $\Delta=12$.

Pre-training paradigm

The paper proposes a curriculum that converts a generic video generator into a policy initialization: (i) initialize from Wan 2.2 trained on web video; (ii) embodied data pre-training on ~10,000 hours combining real robot videos (Agibot, RDT, RoboMind, ATARA, Open X-Embodiment, DROID) and egocentric human videos (EgoDex, Ego4D); (iii) post-training on target-robot trajectories pairing images, language, and actions. The ablation shows these stages are complementary: training from scratch yields SR 0.45, video-model initialization alone 0.57, embodied pre-training alone 0.73, and both combined 0.83. Scaling embodied pre-training data improves real-world success monotonically from 57% to 83%, suggesting continued gains from broader embodied coverage.

Results

Inference efficiency. On an A100, GigaWorld-Policy runs at 360 ms per inference versus 3,231 ms for Motus — a $9\times$ speedup — while remaining close to $\Delta$0's 225 ms. The strong claim here is that this speedup comes without sacrificing performance: simulation SR is 0.86 vs. Motus's 0.88, while real-world SR is higher (0.83 vs. 0.76).

Simulation. On RoboTwin 2.0 across 50 tasks under domain randomization, GigaWorld-Policy averages 0.87 (clean) / 0.85 (randomized), comparable to Motus (0.89/0.87) and well above X-VLA (0.73/0.73) and $\Delta$1 (0.43/0.44). The paper reports a 95% improvement over $\Delta$2 on RoboTwin 2.0 and a 20% improvement over VLA baselines under comparable speed settings. Notably, GigaWorld-Policy is not uniformly best — X-VLA wins on several tasks (e.g., Hanging Mug, Move Playingcard Away, Open Microwave), indicating residual weaknesses on fine-grained stacking and precision-placement tasks.

Real world. Across four PiPER-arm tasks (Clean the Desk, Scan a QR Code, Sweep up Trash, Stack Bowls), GigaWorld-Policy averages 0.83 versus Motus's 0.76 and $\Delta$3's 0.69. The authors attribute the advantage over slower WAMs partly to closed-loop control quality: lower latency raises the effective action update rate and strengthens error correction.

Data efficiency. GigaWorld-Policy matches the maximum success rate achieved by $\Delta$4 using only 10% of the training demonstrations, an important result given the cost of real-world data collection.

Ablations

Three ablations isolate design choices. Future-frame count: removing video prediction entirely ($\Delta$5) drops success rate substantially (the text reports 0.60 in one table and cites a gain from 0.65 to 0.83 elsewhere — a minor internal inconsistency), with peak performance at moderate horizons ($\Delta$6, SR 0.83); denser prediction yields diminishing returns and can degrade performance. Causal mask: causal self-attention matches unconstrained self-attention on SR (0.83 vs. 0.81) but is what makes video decoding optional at deployment; it also improves generation quality (PSNR 28.41 vs. 27.87, SSIM 0.901 vs. 0.892), which the authors attribute to eliminating leakage that weakens action-conditioned dynamics learning under full self-attention. Pre-training curriculum: as noted above, video-model initialization and embodied pre-training contribute independently and additively.

Limitations and open questions

Several limitations are acknowledged or evident. The evaluation covers a single 6-DoF gripper platform (AgileX PiPER) and four real-world tasks with 20 trials per task, leaving cross-embodiment generalization untested despite the heterogeneous pre-training corpus. The claim that video prediction serves as a useful auxiliary signal rests on a fixed stride and loss weighting ($\Delta$7, $\Delta$8); the sensitivity of results to these hyperparameters is not characterized. The internal inconsistency in reporting the no-video-prediction baseline (0.60 vs. 0.65) makes the exact magnitude of the dynamics-supervision benefit uncertain. Finally, whether the action-only inference path forfeits benefits that test-time video rollout could provide — e.g., planning or verification via predicted futures, as explored in Cosmos-Policy — remains an open question the paper does not address.

Conclusion

GigaWorld-Policy demonstrates that a video-generation backbone can be converted into an efficient robot policy by making future visual dynamics a training-time regularizer rather than an inference-time requirement. The causal masking scheme is the central mechanism: it decouples action decoding from video synthesis, yielding a $\Delta$9 speedup over Motus with a 7% higher real-world success rate, and a 95% improvement over $\{T_s, T_o\}$0 on RoboTwin 2.0, while matching $\{T_s, T_o\}$1's performance with 10% of its demonstration data. The main open questions concern generalization beyond the evaluated embodiment and whether optional video rollouts can be exploited productively at test time.