Continue training
Every simulo run starts training from scratch unless you name a checkpoint
to start from. Name one with --from:
simulo run train.py --num-envs 512 --max-iterations 900 \ --from policy_brisk-heron-4f8k2m:latestlatest is the last checkpoint the job saved; best is the checkpoint with
the highest reward, which may come from an earlier iteration than latest.
simulo policy list shows your policies and which checkpoints each one has.
The new job makes its own policy; the policy you started from never changes.
Evaluate the new policy’s checkpoint before you use it. To check whether more
training helped, pass the new checkpoint with --policy and the checkpoint you
continued from with --compare. Evaluate your
policy shows the commands and report.
Continue training is the full guide: how
--max-iterations counts against the checkpoint, what happens when you
change the task between runs, and what happens when --from is refused.
Choose and validate assets: train a different robot.