Huaicong Fang

@taipain thanks for the solution. It helps to set ' dtype=torch.bfloat16' in meth:torch.autocast, which address the problem of 'NAN' loss and black predition.

> reset > self.task.step(actions) > File "/home/jaho/pythonProject/SeqDex-master/SeqDex/dexteroushandenvs/tasks/hand_base/base_task.py", line 135, in step > self.pre_physics_step(actions) > File "/home/jaho/pythonProject/SeqDex-master/SeqDex/dexteroushandenvs/tasks/block_assembly/allegro_hand_block_assembly_orient.py", line 1712, in pre_physics_step > self.reset_idx(env_ids, goal_env_ids) > File "/home/jaho/pythonProject/SeqDex-master/SeqDex/dexteroushandenvs/tasks/block_assembly/allegro_hand_block_assembly_orient.py", line 1607, in reset_idx...

hi~ I found the problem is triggered by out of the diaplay memory. `PxgCudaDeviceMemoryAllocator fail to allocate memory 67108864 bytes!! Result = 2` It seems to be w.r.t the error...

@cypypccpy hi~ While going through the code, I also came across this part that puzzled me. It seems that in the forward progress of `allegro_hand_block_assembly_orient.py`, the t_value network without training...