#!/usr/bin/env python3 """Offline import using mjlab's Python environment. Usage: /path/to/mjlab/.venv/bin/python app/scripts/import-t1.py /path/to/mjlab Requires BOOSTER_ASSETS_ROOT if booster_assets is not at mjlab's default path. """ import hashlib import json from pathlib import Path import re import sys import xml.etree.ElementTree as ET import onnx source = Path(sys.argv[1]).resolve() sys.path.insert(0, str(source / 'src')) from mjlab.entity import Entity from mjlab.asset_zoo.robots.booster_t1.t1_constants import ( get_t1_robot_cfg, get_t1_xml, T1_ACTION_SCALE, ) # Select the newest training run, then its highest numbered checkpoint. runs = sorted(p for p in (source / 'logs/rsl_rl/t1_flat_velocity').iterdir() if p.is_dir() and list(p.glob('model_*.pt'))) run = runs[-1] checkpoint = max(run.glob('model_*.pt'), key=lambda p: int(p.stem.split('_')[-1])) policy = run / f'{run.name}.onnx' if not policy.exists() or policy.stat().st_mtime < checkpoint.stat().st_mtime: raise RuntimeError(f'Export the latest checkpoint first: {checkpoint}') # Snapshot the export before model compilation: training may still be running. policy_bytes = policy.read_bytes() metadata = {p.key: p.value for p in onnx.load_model_from_string(policy_bytes).metadata_props} names = metadata['joint_names'].split(',') if len(names) != 23 or metadata['observation_names'] != 'base_lin_vel,base_ang_vel,projected_gravity,joint_pos,joint_vel,actions,command': raise RuntimeError('Unexpected T1 policy contract') defaults = dict(zip(names, map(float, metadata['default_joint_pos'].split(',')), strict=True)) target = Path(__file__).resolve().parents[1] / 'public/robots/t1' target.mkdir(parents=True, exist_ok=True) records = [] def copy(src, dst, data=None): data = src.read_bytes() if data is None else data dest = target / dst dest.parent.mkdir(parents=True, exist_ok=True) dest.write_bytes(data) records.append({'source': str(src), 'file': dst, 'bytes': len(data), 'sha256': hashlib.sha256(data).hexdigest()}) # Entity applies the same actuator and collision editors used during training. entity = Entity(get_t1_robot_cfg()) entity.spec.compile() root = ET.fromstring(entity.spec.to_xml()) for tag in ('keyframe', 'size'): for element in root.findall(tag): root.remove(element) ET.SubElement(root, 'option', timestep='0.005', integrator='implicitfast', solver='Newton', cone='pyramidal', iterations='10', ls_iterations='20', ccd_iterations='50') assets = [] xml = get_t1_xml() for mesh in root.findall('./asset/mesh'): filename = mesh.attrib['file'] dst = f'meshes/{filename}' copy(xml.parent / 'meshes' / filename, dst) assets.append({'path': dst, 'vfs': dst}) ET.indent(root) ET.ElementTree(root).write(target / 'model.xml', encoding='unicode') copy(policy, 'policies/walk.onnx', policy_bytes) copy(xml.parents[2] / 'LICENSE', 'LICENSES/booster-assets.txt') copy(source / 'LICENSE', 'LICENSES/mjlab.txt') groups = {'policy': names} actions = [] for name in names: groups[name] = [name] scale, = [value for pattern, value in T1_ACTION_SCALE.items() if re.fullmatch(pattern, name)] actions.append({'type': 'joint_position', 'group': name, 'scale': scale, 'offset': 'default'}) config = { 'schema': 1, 'id': 't1', 'name': 'Booster T1', 'family': 't1', 'adapter': 'generic', 'dimensions': {'height': 1.18}, 'model': 'model.xml', 'assets': assets, 'base': {'type': 'floating', 'root_body': 'trunk', 'freejoint': 'world_joint', 'spawn': {'pos': [0, 0, 0.6811], 'yaw': 0}}, 'sim': {'timestep': 0.005, 'decimation': 4}, 'joints': {'groups': groups, 'defaults': defaults}, 'observations': {'obs': {'terms': [ {'type': 'linear_velocity', 'sensor': 'imu_lin_vel'}, {'type': 'gyro', 'sensor': 'imu_ang_vel'}, {'type': 'projected_gravity', 'body': 'trunk'}, {'type': 'joint_pos_rel', 'group': 'policy'}, {'type': 'joint_vel', 'group': 'policy'}, {'type': 'last_action', 'group': 'policy'}, {'type': 'command', 'size': 3}, ]}}, 'actions': actions, 'actuation': {'type': 'native_position'}, 'commands': {'size': 3, 'limits': [0.6, -0.4, 0.6]}, 'skills': [{'id': 'walk', 'kind': 'primary', 'owns': 'policy', 'controller': {'type': 'onnx', 'policy': 'policies/walk.onnx'}}], 'input': {'bindings': [ {'action': action, 'codes': codes, 'label': label} for action, codes, label in [ ('move_forward', ['KeyW', 'ArrowUp'], 'Forward'), ('move_backward', ['KeyS', 'ArrowDown'], 'Backward'), ('turn_left', ['KeyA', 'ArrowLeft'], 'Turn left'), ('turn_right', ['KeyD', 'ArrowRight'], 'Turn right'), ('reset', ['Space'], 'Reset'), ('pause', ['Escape'], 'Pause / continue'), ('chaseToggle', ['KeyC'], 'Restore chase camera'), ]]}, 'render': {'source': 'mjcf_visual', 'groups': [1], 'camera': {'distance': 2.5}}, 'presentation': {'theme': 'comic'}, 'termination': {'tilt_z': -0.5, 'min_height': 0.42, 'on_fall': 'hold'}, 'license': 'See LICENSES/booster-assets.txt and LICENSES/mjlab.txt; policy is locally trained.', } (target / 'robot.json').write_text(json.dumps(config, indent=2) + '\n') (target / 'source-manifest.json').write_text(json.dumps({ 'checkpoint': str(checkpoint), 'checkpoint_sha256': hashlib.sha256(checkpoint.read_bytes()).hexdigest(), 'model_source': str(xml), 'files': records, 'behaviorReferences': [ {'file': str(p.relative_to(source)), 'sha256': hashlib.sha256(p.read_bytes()).hexdigest()} for p in [source / 'src/mjlab/asset_zoo/robots/booster_t1/t1_constants.py', source / 'src/mjlab/tasks/velocity/config/booster_t1/env_cfgs.py', source / 'src/mjlab/tasks/velocity/velocity_env_cfg.py']], }, indent=2) + '\n') print(f'Imported {checkpoint.name} / {policy.name} into {target}')