tars-mini / sim /cpg_study.py
tuanhqv123's picture
TARS mini: CAD, MuJoCo sim, trained policies and 5-action demo
4b7148d verified
Raw History Blame Contribute Delete
2.06 kB
"""Domain-randomised test of the smooth CPG walk on the cheapest build (24 trials per setting)."""
import json, sys
import numpy as np
from multiprocessing import Pool
from model import Params
from sim import TarsSim
from cpg_gait import cpg_walk
from speed_study import CHEAP, RANGES
def trial(a):
T, As, Ah, seed = a
rng = np.random.default_rng(seed)
kw = dict(CHEAP, **{k: float(rng.uniform(*v)) for k, v in RANGES.items()})
kw["tau_stall_kgcm"] = 1.85 * float(rng.uniform(0.85, 1.15))
s = TarsSim(Params(**kw)); s.run(0.5)
y0, x0, n0, t0, yaw0 = s.d.qpos[1], s.d.qpos[0], len(s.log["t"]), s.d.time, s.log["yaw"][-1]
cpg_walk(s, T=T, cycles=4, As=As, Ah=Ah); s.run(0.8)
fwd = -(s.d.qpos[1] - y0) * 100
tq = s.torque_stats(); sw = max(tq["sL"]["p95"], tq["sR"]["p95"])
tilt = max(s.log["up"][n0:])
walks = fwd >= 3 and tilt <= 15 and s.log["up"][-1] < 5
return dict(T=T, As=As, Ah=Ah, fwd=fwd, dur=s.d.time - t0, tilt=tilt, yaw=s.log["yaw"][-1] - yaw0,
swing_p95=sw, I_peak=s.current_stats()["peak50ms_A"], walks=bool(walks), ok=bool(walks and sw <= 70))
if __name__ == "__main__":
combos = [(T, As, Ah) for T in (1.0, 1.4, 2.0, 2.8) for As in (7.5, 10.0) for Ah in (15.0,)]
with Pool(9) as p:
R = p.map(trial, [(T, As, Ah, 3000 + i) for T, As, Ah in combos for i in range(24)])
json.dump(R, open("../results/cpg_study.json", "w"), indent=1)
for T, As, Ah in combos:
X = [r for r in R if (r["T"], r["As"], r["Ah"]) == (T, As, Ah)]
v = np.array([r["fwd"] / r["dur"] for r in X])
print(f"T {T:3.1f}s swing ±{As:4.1f}% : walks {100*np.mean([r['walks'] for r in X]):5.1f}% in-torque {100*np.mean([r['ok'] for r in X]):5.1f}% "
f"velocity {np.median(v):4.2f} cm/s [{np.percentile(v,10):4.2f}..{np.percentile(v,90):4.2f}] "
f"max tilt {max(r['tilt'] for r in X):5.1f}° yaw {np.median([r['yaw'] for r in X]):5.1f}° swing p95 {np.median([r['swing_p95'] for r in X]):3.0f}% I_peak {max(r['I_peak'] for r in X):.2f}A")