Issues with Replicating AnymalTerrain task with Unitree Go1
mohamedamr13 opened this issue · 0 comments
mohamedamr13 commented
I am trying to train the UniTree Go1 robot on the AnymalTerrain Task on Isaac Sim 2023.1.1. I am using the usd for go1 provided in the Isaac Sim assets. I then followed this tutorial to make the asset instanceable. However, there is an error once training starts in the quat_rotate_inverse method called in the anymal_terrain.py script line 430
self.base_lin_vel = quat_rotate_inverse(self.base_quat, self.base_velocities[:, 0:3])
2024-01-14 14:55:30 [279,995ms] [Error] [omni.physx.plugin] PhysX error: Please select Physics Scene (create one if there is none), navigate to GPU section and increase "Gpu Found Lost Aggregate Pairs Capacity" value to at least 596379372.
More details:
The application needs to increase PxgDynamicsMemoryConfig::foundLostAggregatePairsCapacity buffers to 596379372, otherwise, the simulation will miss interactions, FILE /buildAgent/work/74336105e89c4a74/source/gpubroadphase/src/PxgAABBManager.cpp, LINE 1153
q [1024, 4]
q_w [1024]
../aten/src/ATen/native/cuda/IndexKernel.cu:92: operator(): block: [36,0,0], thread: [64,0,0] Assertion `index >= -sizes[i] && index < sizes[i] && "index out of bounds"` failed.
File "/home/user/roshdim1/.local/share/ov/pkg/isaac_sim-2023.1.0-hotfix.1/exts/omni.isaac.core/omni/isaac/core/utils/torch/rotations.py", line 147, in quat_rotate_inverse
a = v * (2.0 * q_w**2 - 1.0).unsqueeze(-1)
~~~~~ <--- HERE
b = torch.cross(q_vec, v, dim=-1) * q_w.unsqueeze(-1) * 2.0
c = q_vec * torch.bmm(q_vec.view(shape[0], 1, 3), v.view(shape[0], 3, 1)).squeeze(-1) * 2.0
RuntimeError: CUDA error: device-side assert triggered
There also seems to be other issues with the Go1.usd referenced by other users in the past version. #58 and #47
I have also documented some of the other issues I noticed while launching the training script here.