Relationship to P08
This supplement organizes the newly uploaded Diffusion Policy derivations and links to the existing P08 paper note. It does not count the same paper twice. The focus is action distributions, denoising, and physical-control timing.
Conditional denoising
Let be a demonstrated action chunk and the visual/proprioceptive condition. A typical DDPM parameterization is
Training learns a conditional denoiser; inference generates actions from noise. Schedules, prediction parameterizations, and samplers differ. One simplified recurrence is not every implementation. The noise index belongs to internal generation time, not physical robot time.
Modes and temporal consistency
Expressive distributions still depend on demonstration coverage, conditioning, and optimization. Joint chunk generation models temporal correlation, but smooth predicted sequences do not guarantee closed-loop stability, contact feasibility, or joint constraints.
Receding-horizon systems predict future actions and execute before observing again. Larger reduces query frequency while extending the interval without fresh visual feedback. More denoising steps change generation quality and compute cost, a separate tradeoff.
Variables needed for reproduction
Record observation length, prediction horizon, executed horizon, denoising iterations, camera timestamps, inference latency, and action units. Absolute versus incremental commands, normalization, actuator limits, and low-level control rate also matter.
Dynamic-target comparisons should retain the same policy and execution budget while measuring observation age and synchronous/asynchronous scheduling. Complete original derivations remain available as source downloads; this edition adds no unverified success-rate or hardware-speed numbers.
Related reading: behavior cloning and bins, π₀, RTC, and DynamicVLA.