SeedPolicy: Reversing the Horizon Scaling Paradox in Robotic Diffusion Policies

Executive Summary

TL;DR: SeedPolicy introduces a breakthrough temporal modeling module called Self-Evolving Gated Attention (SEGA). It transforms the standard Diffusion Policy (DP) from a fixed-window observer into a recurrent, context-aware agent. By utilizing a "Self-Evolving Gate" derived from attention maps, the model filters out visual noise and maintains a "memory" of task progress, allowing it to excel in long-horizon tasks where traditional DPs fail.

Background Positioning: This work targets a known but unsolved "glitch" in the Diffusion Policy literature: the fact that adding more history can sometimes make the robot perform worse. SeedPolicy is a structural SOTA upgrade that solves this bottleneck without the massive compute overhead of long-window Transformers.

Problem: The Horizon Scaling Paradox

In robotic manipulation, more information should lead to better decisions. However, standard Diffusion Policies treat history as a simple stack of frames. As the number of frames grows:

1. Quadratic Complexity: Attention costs explode.
2. State Aliasing: Different stages of a task (e.g., before vs. after picking a block) look visually identical, causing the robot to "freeze" or "loop."
3. Temporal Noise: Background shifts and small movements dilute the important semantic cues needed for the task.

Methodology: Self-Evolving Gated Attention (SEGA)

The core innovation is SEGA, which replaces frame-stacking with a recurrent-style latent state update. Instead of looking at $N$ frames, the model maintains a fixed-size Latent State that evolves over time.

1. Dual-Stream Interaction

As shown in the architecture, SEGA operates two parallel paths:

• State Update (Upper Stream): Takes the current observation, extracts new info via Multi-Head Self-Attention (MSA), and merges it into the historical state.
• State Retrieval (Lower Stream): Uses the "memory" in the state to enhance the current image features, effectively providing the robot with "contextual eyeglasses."

2. The Self-Evolving Gate (SEG)

How does the model know which frames are important? The authors use the Cross-Attention maps themselves as a gating signal. If the attention score is low, the gate closes, preventing noisy or irrelevant background shifts from "polluting" the robot's memory.

Experiments: Breaking the Bottleneck

Tested on the RoboTwin 2.0 benchmark (50 tasks), SeedPolicy showed its true strength as tasks became longer.

• Short Tasks: Gains are modest because the fixed window is "enough."
• Long Tasks: SeedPolicy widens the gap significantly (+21.9%), while the baseline DP collapses because it "forgets" where it is in the multi-step process.

Real-World Robustness

In real-world tests (Dexmal DOS-W1 robot), SeedPolicy fixed the "execution stagnation" issue. In a "Bottle Handover" task, standard DP often froze because the visual state mid-task looked too much like the start state. SeedPolicy used its evolving latent state to realize the task was already in progress, achieving a 56% success rate vs 15% for the baseline.

Critical Analysis & Conclusion

Takeaway: SeedPolicy proves that temporal modeling isn't just about "stacking more frames." It's about efficiently compressing that history into a meaningful state while dynamically ignoring noise.

Limitations: While SeedPolicy is efficient, it still struggles in "Hard" randomized settings compared to massive Foundation Models (VLAs) because it lacks the internet-scale visual prior. However, replacing the temporal backbone of those VLAs with SEGA could be a very promising future direction.

Future Outlook: We expect this "gated recurrent attention" approach to become a standard design pattern for lightweight but "smart" edge robotics where 1.2B parameter models are too heavy to run in real-time.