at237299966/ode_policy_microwave_deploy_metrics

ODE Policy vs. FM-DiT vs. Diffusion Policy on the R1 Lite — latency and action smoothness ("microwave")

Real-robot deployment logs, the analysis script and the resulting tables comparing the inference latency and the action smoothness of three policies trained on the same two-arm demonstrations (at237299966/r1lite_microwave, 34 episodes, 15 Hz). Task success rates were scored separately and are not part of this package.

Setting	Checkpoint	Serving
ODE Policy	ODE_policy_microwave 300.ckpt	RK4, 8 steps per chunk, each chunk integrated on from the last one
FM-DiT	fm_dit_microwave 300.ckpt	10 midpoint steps
DP (DDIM-16)	dp_microwave 300.ckpt	--sampler ddim --num_inference_steps 16
DP (DDPM-100)	same weights	the trained 100-step DDPM sampler

All four re-plan every 8 ticks (--replan_every 8) on an RTX 5090 host; the robot-side client (main_r1lite_closedloop.py, 15 Hz, executes row 0 of every reply) reaches the host over WiFi.

Contents (ode_policy_microwave_deploy_metrics.zip)

README.md
results/
  RESULTS_zh.md      the five tables with their analysis (Chinese): data cleaning, policy compute, end-to-end
                     timing, command smoothness, measured smoothness; the filtering rule; metric definitions
  summary.md         the same tables as the script prints them (English), with p-values
  episodes.csv       every metric per episode (and per demonstration), with whether the episode is used
  robustness/        summary_net40ms.md, summary_net66.7ms.md: the tables with the transport threshold at 40 / 66.7 ms
raw/
  server/            serve_policy.py --log_ticks: <stem>[_NNN].jsonl per episode (+ _summary.json, .png)
  client/            main_r1lite_closedloop.py --log-every 1: eval_<stem>.jsonl (steps), times_<stem>.jsonl (per-episode summaries)
scripts/
  deploy_metrics.py  the analysis (same file as neural_ode/real/deploy_metrics.py in the ODE_Policy repo)
  run.sh             fetches the demonstrations' parquet tables and recomputes results/ into results_rerun/
                     (the 40 / 66.7 ms variants into results_rerun/net40ms/, results_rerun/net66.7ms/)

Stems: mw_ode300, mw_fmdit300, mw_dp300 (DDIM-16), mw_dp300_ddpm100.

Reproduce

unzip ode_policy_microwave_deploy_metrics.zip && cd ode_policy_microwave_deploy_metrics
bash scripts/run.sh        # python >= 3.10 with numpy, scipy, pyarrow; the `hf` CLI for the demos
diff results_rerun/summary.md results/summary.md

Produced with Python 3.11, numpy 2.4.3, scipy 1.17.1, pyarrow 24.0.0.

Notes on the data

• Client step k is server tick k + 2 (the client sends two warm-up queries before its loop); the script checks the pairing on the commanded actions of every episode.
• Dropped: two ODE episodes whose label (t02, t09) was re-run (the earlier, aborted attempt); a DP server connection with no client record (warm-up only).
• Network disturbances are removed per control step, by a rule that only looks at the network: a step whose transport time (client round trip − server compute) exceeds 50 ms, the steps in which the client catches up on the delay this caused (a slow policy's own catch-up, e.g. DDPM-100's, is kept), and the 8 steps after. The rest is cut into whole chunks, stretches of at least 4 chunks; an episode is used with at least 10 s of clean data. The first 4 s of DP episode t05 are also removed (another process was using the host GPU).
• This removes 57.8 % of the ODE steps (the 2026-10-15 session had the worst WiFi: transport p99 2.5 s, longest round trip 7 s), 34.8 % of FM-DiT's, 17.4 % of DP-DDIM's and 40.3 % of DDPM-100's. Episodes used: 8 / 16 / 6 / 2; DDPM-100's numbers are indicative only. At 40 ms or 66.7 ms the ratios and significance barely change (results/robustness/); at 40 ms no DDPM-100 episode is left.
• Measured joints are resampled on the client's observation times (log time − round trip − control call); the server's receive times would mix inference time and network jitter into the time axis.
• Logs contain private LAN addresses of the host and the robot computer, joint states and commands; no images.