prlq's picture
pi05_v3/code: openpi patch (vs upstream 215abfb) + server + instructions for running the pi0.5 checkpoint elsewhere; REPORT.md 3.4 / 2.4 updated
cfb943d verified
|
Raw History Blame Contribute Delete
1.74 kB
# Running the π0.5 MobileManipBench checkpoint (`pi05_v3/`) outside our cluster
The checkpoint needs a few additions to upstream openpi (data config `pi05_mmb_v3`, the shared definition
`mmb_canonical_action.py`, the per-robot normalization `mmb_policy.py`, and the normalization statistics).
`openpi_mmb.patch` contains all of them; applying it to upstream commit `215abfb` reproduces our training code exactly.
```bash
git clone https://github.com/Physical-Intelligence/openpi && cd openpi
git checkout 215abfb
git apply /path/to/openpi_mmb.patch # + src/openpi/policies/mmb_canonical_action.py, mmb_policy.py,
# config pi05_mmb_v3, assets/pi05_mmb_v3/mmb/v2t/*.json, ...
# set up the environment as in openpi's README (uv sync ...)
# checkpoint (params + assets, 12.4 GB)
huggingface-cli download prlq/mobilemanipbench-data --repo-type dataset --include "pi05_v3/80000/*" --local-dir ckpt
# policy server with the shared interface (REPORT.md §2.4); compiles once (~30 s) before accepting clients
OPENPI_SKIP_IMAGE_KEYS=right_wrist_0_rgb uv run python /path/to/serve_mmb_policy.py --checkpoint ckpt/pi05_v3/80000 --port 8000
```
- `OPENPI_SKIP_IMAGE_KEYS=right_wrist_0_rgb` skips the always-empty third camera slot (identical results, faster);
it was set during training.
- Without a server, the same objects can be used in-process: `serve_mmb_policy.load_policy("pi05_mmb_v3", ckpt).infer(obs)`
with the request dict of REPORT.md §2.4 returns `{"actions": [10, 14]}`.
- `test_server_client.py` is the client-side test we ran (needs the merged dataset; adapt the paths).
- Tested: ~70 ms per request on one B300; first request after start-up ~30 s (compile) unless warmed up.