💡Abstract

Vision-Language-Action (VLA) models have recently demonstrated strong generalization capabilities in robotic manipulation. Some existing VLA approaches attempt to improve action accuracy by explicitly generating linguistic reasoning traces or future visual observations before action execution. However, explicit reasoning typically incurs non-negligible inference latency, which constrains the temporal resolution required for robotic manipulation. Moreover, such reasoning is confined to the linguistic space, imposing a representational bottleneck that struggles to faithfully capture ineffable physical attributes. To mitigate these limitations, we propose LaST$_0$, a framework that enables efficient reasoning before acting through a Latent Spatio-Temporal Chain-of-Thought (CoT), capturing fine-grained physical and robotic dynamics that are often difficult to verbalize. Specifically, we introduce a token-efficient latent CoT space that models future visual dynamics, 3D structural information, and robot proprioceptive states, and further extends these representations across time to enable temporally consistent implicit reasoning trajectories. Furthermore, LaST$_0$ adopts a dual-system architecture implemented via a Mixture-of-Transformers design, where a reasoning expert conducts low-frequency latent inference and an acting expert generates high-frequency actions conditioned on robotics-oriented latent representations. To facilitate coordination, LaST$_0$ is trained with heterogeneous operation frequencies, enabling adaptive switching between reasoning and action inference rates during deployment. Across ten simulated and six real-world manipulation tasks, LaST$_0$ improves mean success rates by 8% and 13% over prior VLA methods, respectively, while achieving substantially faster inference. Moreover, LaST$_0$ demonstrates strong long-horizon robustness, attaining nearly a 5× higher success rate at the final step of a multi-step real-world task.

🌟Demonstrations

(To avoid collisions, a fixed timesleep is used)

📊Simulation Experiments

LaST$_0$ achieves a mean success rate of 82% across 10 RLBench manipulation tasks, substantially outperforming all prior baselines. These improvements primarily stem from the latent reasoning representations produced by the reasoning expert, which condition the action expert with compact latent states encoding future visual dynamics, 3D spatial structure, and robot proprioceptive information, enabling more stable and temporally coherent action generation.

📊Real-world Experiments

As shown in Table II, LaST$_0$ achieves the best overall performance on real-world manipulation tasks, with a mean success rate of 72%, substantially outperforming SpatialVLA (39%), $\pi_{0.5}$ (59%), and CoT-VLA (53%). LaST$_0$ consistently delivers strong gains across a diverse set of tasks, particularly those requiring precise spatial reasoning and temporally coherent control.

LaST$_0$ also maintains markedly higher success rates across all stages (0.66 ->0.47 -> 0.33) compared to $\pi_{0.5}$ (0.47 -> 0.20 -> 0.07), with the performance gap widening as the horizon increases.