Training & Inference
Selecting GR00T reduces the number of parameter fields; is this an interface fault
No; this is expected behaviour. When GR00T N1.5 is selected, only batch size, steps, and save frequency are displayed, as the remaining four fields (seed, num workers, eval frequency, and log frequency) are not received by GR00T's training process.
The fields displayed are the fields that take effect. See Training Parameters.
Values are overwritten when the policy is switched after parameter adjustment
This is expected behaviour. The parameter set for the previous policy is not applicable to the new one, and the platform therefore populates the new policy's defaults.
The correct sequence is to determine the policy first and adjust the parameters subsequently.
GR00T's loss declines by a limited magnitude; does this indicate that learning has not occurred
Generally not. GR00T fine-tunes a pretrained model and possesses considerable capability from the outset, so its loss begins at a lower value with a smaller magnitude of decline; ACT learns from an initial state and its variation over the run is more pronounced.
Loss values are not comparable between policies. Comparison of the two should be based on actual inference performance. See Training Monitoring.
GPU (TensorRT) is not offered on the training page
This is expected. GPU (TensorRT) appears only on the inference page, as it belongs to the deployment stage. The training page lists only GPU (CUDA) and CPU; where the host is not equipped with an NVIDIA GPU, GPU (OpenVINO) and CPU are listed instead.
Training terminates during execution due to insufficient memory
The batch size should be reduced first, as it is the parameter with the most direct effect on GPU memory consumption.
It should also be confirmed that a second training run is not being executed concurrently.
Loss does not decline; would extending the training duration improve the result
No. The cause generally lies in data quality or policy selection, and extending the training duration will not alter the result.
Resume training constitutes the extension of an existing run rather than its correction. Where the data or parameters are to be changed, a new training run should be created. See circumstances in which resume training is not applicable.
May a model trained with ACT be converted to TensorRT
No. Each conversion path accepts only one policy: ACT corresponds to CPU, OpenVINO, and CUDA; GR00T corresponds to TensorRT and CUDA. No intersection exists between them.
This is the reason policy selection simultaneously determines the deployment hardware. See Policy Selection.
The policy path field is cleared after switching the inference device
This is expected behaviour. Different devices require different model formats, and retaining the path from the previous device would result in a load failure, so the platform clears the field on switching. The path may simply be selected again.
The arm's motion corresponds to the demonstration but a persistent alignment error remains
In most cases this is a viewpoint issue: the model lacks the information required to determine precise position, with the placement of the wrist camera being the principal factor.
The inference page also provides a failure diagnosis reference table permitting the symptom to be traced to the corresponding stage.