ฯ0.5 two-skill student, run C0: pick up the spoon, scoop the beans (YAM Ultra 2 + Sharpa Wave)
A ฯ0.5 policy fine-tuned with openpi for a YAM Ultra 2 arm with a Sharpa Wave hand (28 joints): "pick up the spoon", then "scoop the beans". Trained in simulation only (Isaac Lab); it has not been run on the real robot yet.
C0 continues the round's run-1 checkpoint (step 5999) for 3000 more steps on the round's planner demonstrations alone,
with a learning rate warming up to 2.5e-5 and decaying to 5e-6. It was trained as the control for run A
(Superstar111/pi05-yam-sharpa-spoon-scoop-runA, the same continuation with 328 relabelled scoops added).
In simulation, on 25 randomised validation scenes with 2 of the 10 predicted rows executed per query, it carried beans out of the tray in 14 and 12 of 25 scenes in two samples (run 1, which it starts from: 7 and 8; run A: 14 and 14). On 10 nominal layouts: 7 of 10 at 2 rows per query and 9 of 10 at 5 rows. On the validation scenes at 5 rows: 15 and 21 of 25 in two samples (run A: 19 and 16; run 1: 14 and 18). The relabelled scoops of run A add nothing measurable at either setting; executing 5 rows per query matters most.
Files
checkpoint/: the openpi checkpoint (params/, and the norm stats inassets/local/main_hcw_nom_n300/).bundle/: the openpi source at the trained commit (openpi_src.tar.gz), its tokenizer (openpi_data.tar.gz),serve_vla.sh(serves the checkpoint on a websocket with openpi'sscripts/serve_policy.py),golden.pyandgolden.npz(recorded queries and answers of a fresh server, for a known-answer check), the sha256 lists of both folders, andDEPLOY_INFO.tsv(run, step, configpi05_maggie_2skill_hcw_nojit, views).
Use
hf download Superstar111/pi05-yam-sharpa-spoon-scoop-c0 --local-dir vla_c0
cd vla_c0
(cd checkpoint && sha256sum -c ../bundle/CHECKPOINT_SHA256SUMS)
(cd bundle && sha256sum -c --ignore-missing BUNDLE_SHA256SUMS)
tar -xzf bundle/openpi_src.tar.gz && tar -xzf bundle/openpi_data.tar.gz && cp bundle/serve_vla.sh bundle/golden.py .
./serve_vla.sh 8000 # needs uv, an NVIDIA GPU with about 10 GB free; the first run builds openpi's environment
cd openpi && uv run python ../golden.py check ../bundle/golden.npz # before anything else queries the server
Request: observation/image (head camera), observation/wrist_image (right wrist camera), observation/image2
(camera at the table's end), each uint8 (160, 284, 3) after a JPEG quality-90 round trip; observation/state (the 28
measured joint angles, rad); observation/previous_command_rad (the 28 joint targets last accepted); prompt
(pick up the spoon or scoop the beans). Answer: actions (10, 28) at 10 Hz: the six arm columns are offsets from
the last accepted target at query time, the 22 hand columns are absolute.
Hand mount
Trained with the hand on its original mount. On the rig's turned mount (one screw position), joint 6 reads 90 degrees
higher for the same hand pose: give the model joint 6 minus 90 degrees in observation/state and
observation/previous_command_rad; its arm outputs are offsets and need no change.
The base model is openpi's pi05_base by Physical Intelligence; see openpi for its terms.