cyy/RoboGauge
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

v0.1.16; Almost finish StressPipeline

Browse Source
This commit is contained in:
wty-yy
2025-12-26 23:51:43 +08:00
parent 65490fe591
commit dad65a4e27
11 changed files with 125 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
CMD.md
View File

@@ -32,7 +32,16 @@ python robogauge/scripts/run.py \
# Stress Pipeline
```bash
python robogauge/scripts/run.py \
--task go2_moe \
--experiment-name debug \
--stress-benchmark \
--stress-terrain-names flat slope stairs_up stairs_down wave \
--stress-num-processes 2 \
--num-processes 3 \
--seeds 0 1 2 \
--frictions 0.5 1.0 1.5 2.0 2.5 \
--headless
```
# Radar/Bar Plot

6
UPDATE.md
View File

@@ -1,8 +1,10 @@
# UPDATE
## 20251226
### v0.1.16
Bugs: level_results会存储到MultiPipeline下, 因为MultiPipeline不是子进程启动的, 会覆盖LevelPipeline的Logger冲突
1. 基本完成StressPipeline, 继续调试完毕
1. 基本完成StressPipeline, 但是绘制进度信息还有问题
2. wave地形的穿模判定非常容易触发, 加入最多穿模判定重启次数为1次, 超过该次数后穿模就不再自动重启了
FIX Bugs: (level_results会存储到MultiPipeline下, 因为MultiPipeline不是子进程启动的, 会覆盖LevelPipeline的Logger冲突) 通过创建新的Logger实现
## 20251225
### v0.1.15
1. 统一terrain的大小为10x10m, 某一边的中点在(0,0,0), 对全部的带等级的terrain都加上边界墙高度10m

8
robogauge/scripts/run.py
View File

@@ -13,8 +13,7 @@ os.environ["OMP_NUM_THREADS"] = "1"
os.environ["MKL_NUM_THREADS"] = "1"
from robogauge.tasks import *
from robogauge.tasks.pipeline.multi_pipeline import MultiPipeline
from robogauge.tasks.pipeline.level_pipeline import LevelPipeline
from robogauge.tasks.pipeline import *
from robogauge.utils.task_register import task_register
from robogauge.utils.helpers import parse_args
@@ -23,7 +22,10 @@ from robogauge.utils.logger import logger
if __name__ == '__main__':
args = parse_args()
if args.multi:
if args.stress_benchmark:
stress_pipeline = StressPipeline(args)
stress_pipeline.run()
elif args.multi:
multi_pipeline = MultiPipeline(args)
multi_pipeline.run()
elif args.search_max_level:

12
robogauge/tasks/gauge/gauge_configs/terrain_levels_config.py
View File

@@ -28,18 +28,6 @@ class TerrainSearchLevelsConfig(Config):
[4.8, 0, 2.092 + 0.1],
[4.8, 0, 2.280 + 0.1],
]
spawns = [
[0.5, 0, 0],
[0.5, 0, 0],
[0.5, 0, 0],
[0.5, 0, 0],
[0.5, 0, 0],
[0.5, 0, 0],
[0.5, 0, 0],
[0.5, 0, 0],
[0.5, 0, 0],
[0.5, 0, 0],
]
class wave:
levels = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]

2
robogauge/tasks/gauge/gauge_configs/wave_gauge_config.py
View File

@@ -19,7 +19,7 @@ class WaveGaugeConfig(BaseGaugeConfig):
'{ROBOGAUGE_ROOT_DIR}/resources/terrains/wave/wave_10.xml',
'{ROBOGAUGE_ROOT_DIR}/resources/terrains/wall/10x10_wall.xml',
]
terrain_spawn_pos = [1.5, 1.25, 1.0] # x y z [m], robot freejoint spawn position on the terrain
terrain_spawn_pos = [1.5, 1.25, 0.0] # x y z [m], robot freejoint spawn position on the terrain
class goals:
class target_pos_velocity: # goal to reach a target position by velocity command

9
robogauge/tasks/pipeline/level_pipeline.py
View File

@@ -15,10 +15,11 @@ from robogauge.utils.logger import Logger
level_logger = Logger() # LevelPipeline logger
class LevelPipeline:
def __init__(self, args):
def __init__(self, args, console_output: bool = True):
self.args = args
self.seeds = args.seeds
level_logger.create(args.experiment_name+'_level', args.run_name)
self.console_output = console_output
level_logger.create(args.experiment_name+'_level', args.run_name, console_output=console_output)
def run(self):
level_logger.info(f"🚀 Starting Level Searcher for '{self.args.experiment_name}'.")
@@ -52,9 +53,9 @@ class LevelPipeline:
def test_level(self, level: int):
level_logger.info(f"🔍 Testing level {level}...")
self.args.level = level
multi_pipeline = MultiPipeline(self.args)
multi_pipeline = MultiPipeline(self.args, console_output=self.console_output)
aggregated_results = multi_pipeline.run()
success_mean = float(aggregated_results['success']['mean'].split(' ')[0])
success_mean = float(aggregated_results['summary']['success']['mean'].split(' ')[0])
all_success = success_mean >= 0.8
if all_success:
level_logger.info(f"✅ Level {level} passed all tests.")

23
robogauge/tasks/pipeline/multi_pipeline.py
View File

@@ -22,6 +22,7 @@ from robogauge.tasks.pipeline.base_pipeline import BasePipeline
from robogauge.utils.task_register import task_register
from robogauge.utils.logger import Logger
from robogauge.utils.process_utils import NoDaemonPool
multi_logger = Logger() # MultiPipeline logger
@@ -62,14 +63,14 @@ def run_single_process(args, data):
return ret
class MultiPipeline:
def __init__(self, args):
def __init__(self, args, console_output: bool = True):
self.args = args
self.seeds = args.seeds
self.frictions = args.frictions
self.base_masses = args.base_masses
self.num_processes = args.num_processes
self.static_info = {}
multi_logger.create(args.experiment_name+'_multi', args.run_name+'_multi')
multi_logger.create(args.experiment_name+'_multi', args.run_name+'_multi', console_output=console_output)
def add_static_info(self, key: str, value):
if key not in self.static_info:
@@ -85,7 +86,7 @@ class MultiPipeline:
ctx = multiprocessing.get_context('spawn')
worker_func = functools.partial(run_single_process, self.args)
results_list = []
with ctx.Pool(processes=self.num_processes) as pool:
with NoDaemonPool(processes=self.num_processes, context=ctx) as pool:
iterator = pool.imap_unordered(worker_func, workers_data)
for results in tqdm(iterator, total=len(workers_data), desc="Evaluation"):
results_list.append(results)
@@ -104,7 +105,7 @@ class MultiPipeline:
""" Process results from all processes and aggregate them. """
multi_logger.info("📊 Aggregating Results from all runs...")
summary = {'success': {}, **self.static_info}
summary = {'success': {}, **self.static_info, 'summary': {}}
finish_msg = (
f"""\n{'='*20} Run Finish Summary {'='*20}\n"""
f"""{'Seed':^10}{'Base Mass':^15}{'Friction':^15}{'Status':^10}\n"""
@@ -133,9 +134,9 @@ class MultiPipeline:
value_collections[metric][mean_name].append(float(mean_value.split(' ')[0]))
for metric, means in value_collections.items():
summary[metric] = {}
summary['summary'][metric] = {}
for mean_name, values in means.items():
summary[metric][mean_name] = f"{float(np.mean(values)):.4f} ± {float(np.std(values)):.4f}"
summary['summary'][metric][mean_name] = f"{float(np.mean(values)):.4f} ± {float(np.std(values)):.4f}"
save_path = multi_logger.log_dir / "aggregated_results.yaml"
with open(save_path, 'w') as file:
@@ -143,9 +144,9 @@ class MultiPipeline:
multi_logger.info("✅ Aggregated execution finished.")
multi_logger.info(f"📁 Aggregated results saved to: {save_path}")
multi_logger.info(
f"""\n{'='*20} Multi-Run Summary {'='*20}\n"""
f"""{yaml.dump(summary, allow_unicode=True)}"""
f"""{'='*60}"""
)
# multi_logger.info(
# f"""\n{'='*20} Multi-Run Summary {'='*20}\n"""
# f"""{yaml.dump(summary, allow_unicode=True)}"""
# f"""{'='*60}"""
# )
return summary

57
robogauge/tasks/pipeline/stress_pipeline.py
View File

@@ -9,12 +9,15 @@
'''
import yaml
import functools
import numpy as np
from tqdm import tqdm
import multiprocessing
from copy import deepcopy
from itertools import product
from collections import defaultdict
from robogauge.utils.logger import Logger
from robogauge.utils.process_utils import NoDaemonPool
from robogauge.tasks.pipeline import MultiPipeline, LevelPipeline
from robogauge.tasks.gauge.gauge_configs.terrain_levels_config import TerrainSearchLevelsConfig
@@ -36,20 +39,22 @@ def run_pipeline(args, data):
level = None # flat terrain
if search:
level, results = LevelPipeline(args).run()
level, results = LevelPipeline(args, console_output=False).run()
if level == 0: # no valid level found
results = {
'success': False,
'results': results,
'data': data,
'level': 0,
}
else:
args.level = level
results = {
'success': True,
'results': MultiPipeline(args).run(),
'results': MultiPipeline(args, console_output=False).run(),
'data': data,
'level': level,
}
return results
@@ -95,19 +100,22 @@ class StressPipeline:
}
if search_max_level:
for friction, base_mass in product(self.args.frictions, self.args.base_masses):
data.update({
now_data = deepcopy(data)
now_data.update({
'friction': friction,
'base_mass': base_mass,
})
workers_data.append(now_data)
else:
workers_data.append(data)
### Run and collect results ###
results_list = []
with ctx.Pool(processes=self.num_processes) as pool:
with NoDaemonPool(processes=self.num_processes, context=ctx) as pool:
iterator = pool.imap_unordered(worker_func, workers_data)
for results in tqdm(iterator, total=len(workers_data), desc="Stress Benchmark"):
results_list.append(results)
self.add_static_info('model_path', results['results']['model_path'])
self.add_static_info('model_path', results['results'].pop('model_path', None))
stress_logger.info("✅ Stress Benchmark Completed.")
stress_results = self.aggregate_results(results_list)
@@ -115,17 +123,48 @@ class StressPipeline:
def aggregate_results(self, all_results):
stress_logger.info("📊 Aggregating Stress Benchmark Results...")
summary = {}
finish_msg = (
f"""\n{'='*20} Stress Benchmark Summary {'='*20}\n"""
f"""{'Terrain Name':^20}{'Base Mass':^15}{'Friction':^15}{'Status':^10}\n"""
f"""{'Seeds':^20}{str(self.seeds)}\n"""
f"""{'Terrain Name':^20}{'Base Mass':^15}{'Friction':^15}{'Max Level':^15}\n"""
)
all_results = sorted(all_results, key=lambda x: (x['data']['terrain_name'], x['data'].get('base_mass', 0), x['data'].get('friction', 0)))
for result in all_results:
terrain_name = result['data']['terrain_name']
base_mass = result['data'].get('base_mass', self.args.base_masses)
friction = result['data'].get('friction', self.args.frictions)
status = "✅" if result['success'] else "❌"
finish_msg += f"{terrain_name:^20}{str(base_mass):^15}{str(friction):^15}{status:^10}\n"
status = f"{result['level']}" if result['success'] else "❌"
finish_msg += f"{terrain_name:^20}{str(base_mass):^15}{str(friction):^15}{status:^15}\n"
finish_msg += f"""{'='*66}"""
stress_logger.info(finish_msg)
if not all_results:
stress_logger.error("No results to aggregate.")
return
summary = {**self.static_info, 'summary': {}}
value_collections = defaultdict(lambda: defaultdict(list))
for result in all_results:
terrain_name = result['data']['terrain_name']
terrain_level = result['level'] # None, 0, 1, ..., 10
key = terrain_name
if terrain_level is not None:
key += f'_{terrain_level}'
key += f'_baseMass{result["data"]["base_mass"]}_friction{result["data"]["friction"]}'
summary[key] = result['results'] if terrain_level != 0 else None
for metric, means in result['results']['summary'].items():
for mean_name, mean_value in means.items():
value_collections[metric][mean_name].append(float(mean_value.split(' ')[0]))
for metric, means in value_collections.items():
summary['summary'][metric] = {}
for mean_name, values in means.items():
summary['summary'][metric][mean_name] = f"{float(np.mean(values)):.4f} ± {float(np.std(values)):.4f}"
save_path = stress_logger.log_dir / "stress_benchmark_results.yaml"
with open(save_path, 'w') as file:
yaml.dump(summary, file, allow_unicode=True, sort_keys=False)
stress_logger.info(f"✅ Stress benchmark aggregated execution finished.")
stress_logger.info(f"📁 Stress benchmark results saved to: {save_path}")
return summary

1
robogauge/tasks/simulator/mujoco_config.py
View File

@@ -52,3 +52,4 @@ class MujocoConfig(Config):
penetration_threshold = -0.035 # [m], if any contact penetration depth is below this threshold, truncate episode
skip_penetration_geoms = ['wall', 'floor'] # Geometries to skip penetration check
skip_self_penetration = True # Whether to check self-penetration
penetration_max_reset_num = 1 # Max number of resets due to penetration per run

13
robogauge/tasks/simulator/mujoco_simulator.py
View File

@@ -44,6 +44,7 @@ class MujocoSimulator:
self.n_step = 0
self.sim_time = 0.0
self.target_pos = None
self.penetration_reset_count = 0
def load(
self,
@@ -158,6 +159,7 @@ class MujocoSimulator:
self._pause = False
self.n_step = 0
self.sim_time = 0.0
self.penetration_reset_count = 0
self.load_dof_limits()
self.preload_sensors()
@@ -269,6 +271,8 @@ class MujocoSimulator:
return sim_data
def check_penetration(self, threshold: float = -0.02):
if self.penetration_reset_count >= self.cfg.truncation.penetration_max_reset_num:
return False, None, None, None
for i in range(self.mj_data.ncon):
contact = self.mj_data.contact[i]
if contact.dist < threshold:
@@ -285,16 +289,17 @@ class MujocoSimulator:
is_penetrated, geom1, geom2, dist = self.check_penetration(self.cfg.truncation.penetration_threshold)
if is_penetrated:
flat = True
is_err = True
if self.cfg.truncation.skip_penetration_geoms is not None and (
any(skip_geom in geom1.lower() for skip_geom in self.cfg.truncation.skip_penetration_geoms) or
any(skip_geom in geom2.lower() for skip_geom in self.cfg.truncation.skip_penetration_geoms)
):
flat = False
is_err = False
if self.cfg.truncation.skip_self_penetration:
if geom1.split('/')[0] == geom2.split('/')[0]:
flat = False
if flat:
is_err = False
if is_err:
self.penetration_reset_count += 1
raise RuntimeError(f"[Penetration Error] Episode truncated: Penetration ({geom1} <-> {geom2}), distance: {dist}")
def reset(self):

29
robogauge/utils/process_utils.py Normal file
View File

@@ -0,0 +1,29 @@
# -*- coding: utf-8 -*-
'''
@File : process_utils.py
@Time : 2025/12/26 21:59:53
@Author : wty-yy
@Version : 1.0
@Blog : https://wty-yy.github.io/
@Desc : No Daemon Pool for Multiprocessing
'''
import multiprocessing
import multiprocessing.pool
class NoDaemonProcess(multiprocessing.Process):
@property
def daemon(self):
return False
@daemon.setter
def daemon(self, value):
pass
class NoDaemonPool(multiprocessing.pool.Pool):
def __init__(self, *args, **kwargs):
""" context=ctx in kwargs is required """
super(NoDaemonPool, self).__init__(*args, **kwargs)
def Process(self, *args, **kwds):
proc = super(NoDaemonPool, self).Process(*args, **kwds)
proc.__class__ = NoDaemonProcess
return proc