TIDAL: Temporally Interleaved Diffusion and Action Loop for High-Frequency VLA Control

TL;DR

The paper addresses the latency bottleneck of large Vision-Language-Action models in dynamic environments by introducing TIDAL, a hierarchical dual-frequency framework that decouples semantic reasoning from high-frequency actuation. It implements a macro-loop for semantic caching and a micro-loop for interleaved single-step flow execution, enabling ~9 Hz updates on edge hardware while maintaining a fixed backbone budget. Through temporally misaligned training and a Differential Motion Predictor, TIDAL compensates for stale semantics and velocity-insensitive perception, achieving a 2x gain in dynamic interception and a 4x increase in feedback frequency. The approach preserves broad VLA capabilities, demonstrates robustness under non-paused inference, and extends the usable semantic horizon beyond the native action chunk, indicating strong potential for real-world dynamic manipulation and sim-to-real transfer.

Abstract

Large-scale Vision-Language-Action (VLA) models offer semantic generalization but suffer from high inference latency, limiting them to low-frequency batch-and-execute paradigm. This frequency mismatch creates an execution blind spot, causing failures in dynamic environments where targets move during the open-loop execution window. We propose TIDAL (Temporally Interleaved Diffusion and Action Loop), a hierarchical framework that decouples semantic reasoning from high-frequency actuation. TIDAL operates as a backbone-agnostic module for diffusion-based VLAs, using a dual-frequency architecture to redistribute the computational budget. Specifically, a low-frequency macro-intent loop caches semantic embeddings, while a high-frequency micro-control loop interleaves single-step flow integration with execution. This design enables approximately 9 Hz control updates on edge hardware (vs. approximately 2.4 Hz baselines) without increasing marginal overhead. To handle the resulting latency shift, we introduce a temporally misaligned training strategy where the policy learns predictive compensation using stale semantic intent alongside real-time proprioception. Additionally, we address the insensitivity of static vision encoders to velocity by incorporating a differential motion predictor. TIDAL is architectural, making it orthogonal to system-level optimizations. Experiments show a 2x performance gain over open-loop baselines in dynamic interception tasks. Despite a marginal regression in static success rates, our approach yields a 4x increase in feedback frequency and extends the effective horizon of semantic embeddings beyond the native action chunk size. Under non-paused inference protocols, TIDAL remains robust where standard baselines fail due to latency.

Figures (3)

• Figure 1: Dynamic Interception Task Execution. A complete rollout where TIDAL (1) intercepts a moving can, (2) places it into the drawer, and (3) closes the drawer. Unlike open-loop baselines, TIDAL's high-frequency micro-control loop enables the robot to reactively track the moving target and robustly complete the long-horizon sequence.
• Figure 2: Comparison of Inference Paradigms in Dynamic Interception. (Top) Standard Baseline: Under the batch-and-execute model, the system performs a blocking inference at the start. It generates a fixed 16-step action chunk conditioned on the static initial state $s_0$. The resulting open-loop execution forces the robot to follow a stale trajectory, missing the moving target. (Bottom) TIDAL (Ours): Our Hierarchical Dual-Frequency Architecture. The macro-loop queries the VLM once to generate a Cached Intent $\mathcal{E}$ (grey bar), decoupling expensive reasoning from high-frequency actuation. The micro-loop performs interleaved flow integration. At each update, the DiT combines the frozen intent with the real-time fused state $\tilde{s}_t$ to generate a fresh trajectory. The policy executes only the immediate chunk ($N=4$) and discards the tail (fading blocks). This sliding-window execution refreshes the control signal at $\sim$9 Hz, enabling the robot to robustly intercept dynamic objects.
• Figure 3: Temporally Misaligned Training Strategy. We decouple the timestamps of visual and physical inputs to handle latency between the macro-Loop and micro-Loop. Latency Injection (Vertical Axis): A single frozen semantic intent (encoded from image $I_0$) is reused across multiple latency stages ($k=0, \dots, 3$). Training pairs this stale intent with progressively later proprioceptive states ($s_{0}, s_{4}, s_{8}, s_{12}$), enabling the policy to learn predictive compensation for VLM lag. Optimization Objectives: We use a Horizon-Weighted Loss (Dark vs. Light Green bars) with weight $w=2.0$ for the immediate execution chunk ($N=4$) to to emphasize execution stability. The Time-Biased Sampling inset ($t \sim P(t)$) shows that we bias flow timestep sampling towards the noise source ($t \approx 0$) to optimize the vector field for single-step integration.