openwam-policy

OpenWAM-alpha world-action model run as a policy (end-effector actions), RoboDojo’s own checkpoint, through XPolicyLab’s OpenWAM adapter

• Kind: policy
• Protocol: xpolicylab
• Default port: 19010
• VRAM: 15.3 GB
• Load time: about 240 s
• Skills that use a policy: vla
• Configurations that name it: –

What the model is told

A world-action model trained on whole RoboDojo tasks: each demonstration starts at the start pose and finishes the task, labelled with the task’s own instruction; it predicts the end-effector motion of both arms. It does best in exactly that setting: from the untouched start pose, with the task’s own instruction, left to run. It cannot pick one object out of several by reading what identifies it: a name, a number or label written on it, or a picture shown as the target. Signs it has stopped helping: the arms stop moving while objects are not where the task needs them; the same reach is repeated; an object is knocked over or off the table.

Weights

harness fetch openwam-policy --deploy <machine>

Directory	Source	From
<weights>/openwam	hf_model	OpenWAM/OpenWAM-Alpha-Sim-RoboDojo

Notes

The same checkpoint as the openwam world-model service; run one of the two per GPU. Its actions are link6 poses for both arms (action_type ee), which run_policy solves by IK every step, as RoboDojo’s client does. The description is generic until OpenWAM-alpha alone has run on development layouts (configs/experiments/policy-alone-dev.toml) and scripts/policy_card.py has sorted its scores into kinds of manipulation.

Manifest

kind = "policy"
summary = "OpenWAM-alpha world-action model run as a policy (end-effector actions), RoboDojo's own checkpoint, through XPolicyLab's OpenWAM adapter"
protocol = "xpolicylab"
port = 19010
# XPolicyLab's server for the adapter would load the checkpoint whole on the GPU with the
# pinned numpy; this one loads it as the openwam world-model service does (text encoder on
# the CPU, the stats re-saved under numpy 1) and speaks XPolicyLab's websocket protocol.
command = ["{python}", "-u", "services/servers/serve_openwam_policy.py", "--ckpt-dir", "{weights}/openwam",
           "--port", "{port}"]
env = { HARNESS_ROBODOJO_ROOT = "{robodojo}" }
# Measured for the openwam world-model service, loaded the same way: 15.3 GB.
vram_gb = 15.3
load_seconds = 240
needs_files = ["{weights}/openwam/config.yaml", "{xpolicylab}/policy/OpenWAM/model.py"]
notes = """The same checkpoint as the openwam world-model service; run one of the two per GPU. Its actions are link6 \
poses for both arms (action_type ee), which run_policy solves by IK every step, as RoboDojo's client does. The \
description is generic until OpenWAM-alpha alone has run on development layouts (configs/experiments/\
policy-alone-dev.toml) and scripts/policy_card.py has sorted its scores into kinds of manipulation."""

[client]
url = "{url}"
description = """A world-action model trained on whole RoboDojo tasks: each demonstration starts at the start pose \
and finishes the task, labelled with the task's own instruction; it predicts the end-effector motion of both arms. \
It does best in exactly that setting: from the untouched start pose, with the task's own instruction, left to run. \
It cannot pick one object out of several by reading what identifies it: a name, a number or label written on it, \
or a picture shown as the target. Signs it has stopped helping: the arms stop moving while objects are not where the \
task needs them; the same reach is repeated; an object is knocked over or off the table."""

# The adapter's protocol (what run_policy sends and gets: ee actions, 16 values). ckpt_name
# and the overrides apply only when XPolicyLab's own server is used (no command above).
[xpolicylab]
policy = "OpenWAM"
ckpt_name = "{weights}/openwam"
action_type = "ee"
action_dim = 16

[xpolicylab.overrides]
ckpt_dir = "{weights}/openwam"
# run_policy asks for one environment's chunk at a time.
eval_batch = "false"

[[weights]]
dir = "openwam"
hf_model = "OpenWAM/OpenWAM-Alpha-Sim-RoboDojo"