1  Hardware Efficiency & Edge Deployment Constraints

Deploying generative continuous control policies (50Hz) to resource-constrained edge robotic hardware (e.g., Jetson Orin) is computationally non-trivial. Standard Denoising Diffusion Probabilistic Models (DDPM) require deep iterative sampling, introducing latency that is fundamentally incompatible with reactive parkour or dynamic manipulation.

This chapter systematically evaluates the empirical edge performance of our 1-step Rectified Flow architecture compared to standard DDPM baselines.

1.1 System Footprint: VRAM and Training Throughput

Before analyzing deployment latency, we quantify the computational efficiency of the dual-objective distillation pipeline during offline pretraining.

1.2 The Generative Latency Bottleneck: \(O(1)\) vs \(O(T)\)

A critical mathematical claim of this research is that Flow Matching—using a single-step Euler integration along a learned straight-line vector field—outperforms traditional 20-step DDPMs in reactive robotic environments without sacrificing the structural integrity of the action manifold.

1.2.1 Statistical Evaluation: Hardware Feasibility

    Wilcoxon rank sum exact test

data:  latency_ms by model_type
W = 36, p-value = 0.002165
alternative hypothesis: true location shift is not equal to 0

Interpretation: The computational overhead of training the dual-objective Vision Transformer and 1D U-Net remains tractable, exhibiting stable Steps-Per-Second (SPS) and negligible VRAM footprints across both locomotion and manipulation domains.

Crucially, the Wilcoxon rank-sum test confirms a statistically significant (\(p < 0.05\)) latency reduction. Traditional 20-step diffusion operates significantly above the acceptable tolerance for 50Hz control loops (requiring \(<20\)ms latency to prevent control saturation). The \(O(1)\) flow-matched policy collapses the generation time, confirming that straight-line ODE flows solve the edge deployment constraint inherent to continuous-time generative visuomotor policies.