tourmii/tfp-dp-button-repro

TFP Small-Scale Reproduction — Diffusion Policy Checkpoints (Timed Button Press)

Checkpoints from a small-scale reproduction of Temporal Flow Policy (arXiv:2610.20524), Diffusion Policy backbone, Timed Button Press task (paper App. D.1 protocol). Full report: tourmii/tfp-repro (bucket README).

Contents

• policy_tfp.pt, policy_gru.pt, policy_reactive.pt — 0.7 M-param state-only Diffusion Policies (128-width transformer denoiser, adaLN-zero, DDPM 100 cosine steps, x0-prediction, DDIM-25 at eval), 40k training steps with DP-style observation-noise augmentation. policy_tfp.pt includes the TFP memory module: 128-channel state, exponential relaxation h = e^{-Δt/τ}⊙h + (1-e^{-Δt/τ})⊙ĥ with candidate and per-channel time-constant heads (softplus, 1e-6 floor, zero-init τ-head), AdaLN conditioning c = c_native + M(h).
• results.json — eval success rates (30 trials × 4 schedules × 3 variants), TFP frozen-weight interventions, event-age probe.
• metrics.jsonl — per-step training losses.
• demos_val.jsonl, demos_meta.json — validation episodes and protocol metadata (also in tourmii/tfp-button-demos-repro).

Headline result

• Reproduced: time-since-event decodes from the frozen TFP memory state at MAE 0.0104 s vs 0.941 s (constant predictor) and 2.307 s (episode clock), n=450 — matching the paper's §4.2/Table 8 conclusion.
• Not reproduced: closed-loop task success is 0.0 for all variants at every schedule (paper: TFP 81.2%, reactive ≈ 0% — we reproduce the reactive failure only). Diagnosis and ruled-out causes in the bucket report.

Citation

arXiv:2610.20524 · Chi et al., Diffusion Policy, arXiv:2303.04137