The trainning problem of BlockAssemblyOrient task

[pp5201314]

Hello. Thank you for your wonderful work. May I ask if it is possible to provide the terminal status file (.pth) and the corresponding checkpoint file obtained from the separate training of each sub-task? The training of the BlockAssemblySearch subtask was successful, but during the 6-hour training period of the -BlockAssemblyOrient task, the terminal state I obtained remained 0 all the time.

[pp5201314]

May I ask if training a sub-task separately will not optimize the grasp-insert-t-value, so the t-value obtained in the Orient task is of no use?

[Ralph-cong]

@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 is used to judge if the intermediate state is as good to be saved. According to paper, it seems that in forward process, the t_value network should not be included.

                        if self.tvalue[i] > 0.6:
                            if self.save_hdf5:
                                if self.use_temporal_tvalue:
                                    self.succ_grp.create_dataset("{}th_success_data".format(self.success_v_count), data=self.t_value_obs_buf[i].cpu().numpy())
                                else:
                                    self.succ_grp.create_dataset("{}th_success_data".format(self.success_v_count), data=self.camera_view_segmentation_target_rot[i].cpu().numpy())
                                
                                self.success_v_count += 1

                            self.saved_digging_ternimal_states_list[object_i][self.saved_digging_ternimal_states_index_list[object_i]:self.saved_digging_ternimal_states_index_list[object_i] + 1] = self.saved_searching_ternimal_state[i]
                            self.saved_digging_ternimal_states_index_list[object_i] += 1
                            if self.saved_digging_ternimal_states_index_list[object_i] > 10000:
                                self.saved_digging_ternimal_states_index_list[object_i] = 0