Technical Summary: Learning High-Frequency Continuous Action Chunks in Latent Space

1. Problem Statement

Modern robotic policies increasingly utilize action chunking—predicting temporally extended action sequences rather than single-step commands—to model complex trajectories and long-horizon dependencies. While effective at moderate action frequencies, this approach faces significant challenges when scaled to high frequencies (e.g., 60 Hz).

At high frequencies, policies trained directly in the action space often fail to generate actions that are both temporally smooth and spatially consistent. This results in:

• Jittery trajectories: High-frequency noise and imprecision.
• Stop-and-go motion: Low-frequency discretization forces repeated acceleration and deceleration, causing velocity drops and discontinuous motion.
• Asynchronous inference gaps: Real-time execution requires overlapping policy inference with action execution. However, switching between asynchronously inferred chunks often introduces large discontinuities at chunk boundaries, leading to visible stalls, rollbacks, and degraded execution quality.

Existing alternatives, such as training at low frequencies and interpolating to high frequencies, amplify prediction errors and fail to recover the fine-grained motion structure required for smooth control.

2. Methodology

The authors propose a two-pronged approach to address these challenges: shifting learning to a latent space and introducing a training-free execution strategy called Reuse-then-Refine (RTR).

2.1. Latent Space Learning via VAE

Instead of predicting high-frequency action chunks directly in the action space, the policy learns to predict in a continuous latent space compressed by a Variational Autoencoder (VAE).

• Encoder: Compresses a high-frequency action chunk $A_t \in \mathbb{R}^{H \times c}$ (where $H$ is the horizon and $c$ is the action dimension) into a lower-dimensional, temporally downsampled latent representation $z \in \mathbb{R}^{h \times d}$ (where $h = H/f$ and $f$ is the downsampling factor).
• Policy Training: The policy $\pi_\theta$ is trained to map observations $o_t$ to latent vectors $z$, rather than raw actions.
• Decoder: A fixed VAE decoder reconstructs the high-frequency action chunk from the predicted latent vector.
• Mechanism: This formulation shifts the learning target from fine-grained, high-frequency command variations to coherent local motion patterns. The KL regularization encourages these patterns to lie on a smoother latent manifold, reducing the burden on function approximation and suppressing spurious high-frequency disturbances.

2.2. Reuse-then-Refine (RTR)

To ensure continuity between adjacent action chunks under asynchronous inference (where inference latency causes misalignment between the predicted chunk and the current robot state), the authors introduce RTR.

• Reuse: Instead of discarding outdated actions from the previous chunk or the newly inferred chunk, RTR reuses the actions executed during the inference window.
• Concatenation: These reused actions are concatenated with the non-outdated portion of the newly predicted action chunk, forming a temporally misaligned intermediate sequence.
• Refine: This concatenated sequence is fed into the VAE encoder to obtain a latent representation, which is then decoded.
• Result: The VAE decoder smooths inconsistencies within the concatenated sequence while preserving alignment with the most recently executed actions, ensuring a seamless transition between chunks without requiring additional policy training.

3. Key Contributions

1. Latent-Space High-Frequency Learning: Demonstrates that learning high-frequency action policies in a continuous latent space (via VAE) yields significantly smoother and more precise trajectories compared to direct action-space learning, particularly for models relying on discrete tokenization (e.g., OpenVLA-OFT).
2. Reuse-then-Refine (RTR): Proposes a training-free strategy to improve chunk-level continuity under asynchronous inference. Unlike prior methods (e.g., RT-C) that condition generation on previous chunks in action space, RTR leverages the VAE to refine mixed sequences of reused and new actions.
3. Empirical Validation: Provides extensive real-world evaluations on three contact-rich tasks (Peel Cucumber, Wipe Vase, Write Board) using three distinct policy architectures (Diffusion Policy, OpenVLA-OFT, PI0.5).

4. Experimental Results

The paper evaluates performance using metrics for precision (Mean Absolute Error/Deviation), smoothness (Jerk), safety (Exceed count of speed limits), and latency.

• Precision and Smoothness: Latent policies consistently outperform their action-space counterparts. For example, on the "Write Board" task, the latent version of OpenVLA-OFT reduced jerk from 5.24 to 0.56 and improved success rates significantly.
• Asynchronous Execution: RTR effectively eliminates the "stalls" and "rollbacks" observed in naive asynchronous execution. In the "Write Board" task with PI0.5, RTR reduced jerk from 4.984 (Original) to 1.754 (Latent+RTR) and eliminated visible rollback artifacts.
• Latency: By enabling smoother execution without repeated acceleration/deceleration and reducing stalls, the proposed method achieves lower end-to-end execution latency compared to low-frequency baselines and other high-frequency alternatives.
• Ablation: The study confirms that moderate temporal compression (e.g., downsampling ratio $f=4$) optimizes the trade-off between precision and smoothness, while aggressive compression degrades performance. It also shows that applying continuity methods like RT-C directly in latent space is ineffective, whereas RTR is specifically designed for this setting.

5. Significance and Claims

The authors claim that their work highlights the importance of representation and execution co-design for high-frequency robotic control.

• Practical Impact: The approach provides a practical pathway for robots to execute complex, contact-rich tasks continuously with fewer pauses and jerky motions in real-world settings.
• Generalizability: The method improves performance across diverse architectures (Diffusion, VLA/Flow-based) and is compatible with existing imitation learning frameworks without requiring retraining of the base policy architecture, only the addition of a VAE and the RTR inference strategy.
• Limitations: The authors note that their experiments are currently constrained to 60 Hz due to sensor sampling rates and that the applicability of RTR to policies trained directly in action space (without a latent policy) remains an open question requiring further investigation.

The paper concludes that learning in latent space combined with RTR enables stable, continuous, and smooth robot execution, advancing the reliability of high-frequency control in physical environments.