bug: 域随机化引起的场景构建卡顿

[humble02]

Pre-flight checklist

• I searched existing issues (open and closed) and found no duplicate.
• I checked the README Troubleshooting section — this is not one of the documented pitfalls.
• The bug does not reproduce in a pure Genesis script without GeneLab. (If unsure, leave unchecked and say so below — simulator bugs belong on the Genesis tracker.)

What happened

unchecked
genesis: rigid_solver.py lines 744-833
# TODO: Simple numpy-based for-loop for now as it is not performance critical
for i_b_, i_b in enumerate(envs_idx):
# Compute the inverted mass matrix efficiently
mass_mat_L_inv = np.eye(self.n_dofs_)
for i_d in range(self.n_dofs_):
for j_d in range(i_d):
mass_mat_L_inv[i_d] -= mass_mat_L[i_d, j_d, i_b] * mass_mat_L_inv[j_d]
mass_mat_inv = (mass_mat_L_inv * mass_mat_D_inv[:, i_b]) @ mass_mat_L_inv.T
# ... per-link Jacobian build + A = jac @ mass_mat_inv @ jac.T ...
# Stop there if not batched
if not batched:
break
当对为 armature/mass/friction 使用域随机化，会开启batch,对每个 env 都跑一遍:先 O(n_dofs²) 求质量矩阵逆,再对每根连杆(~30)组雅可比做矩阵乘计算,在单核 CPU 上串行。导致训练启动非常慢，4096环境需约15～20分钟。
以上为gpt5.6给我的解释。

Steps to reproduce

uv run genelab train Genelab-Velocity-Mattress-Unitree-G1-v0 --num_envs 4096

Environment

genelab 0.4.0
genesis 1.2.1
torch 2.10.0+cu128 cuda 12.8

RL backend

rsl_rl

Simulation backend

GPU (gpu=True)

Logs / traceback

1

Anything else

No response