Tracking episode-level experience across training restarts

[robotmem]

When running GPU-parallelized training on PickCube or PushCube tasks,
restarting a training run (say after adjusting reward shaping or domain
randomization) means losing the episode-level context from the previous
run. Checkpoints save the policy, but not things like "which object
positions were harder" or "reward scale of 10 caused instability."

I've been using a simple wrapper that logs episode outcomes to a local
SQLite file and retrieves relevant past entries on environment reset:

import gymnasium as gym
import mani_skill.envs
from robotmem.gymnasium import RobotMemWrapper

env = gym.make("PushCube-v1", obs_mode="state")
env = RobotMemWrapper(env, collection="maniskill_push")
# auto-logs episodes, retrieves tips on reset

Since ManiSkill uses the standard Gymnasium API, the wrapper doesn't
touch ManiSkill internals. The SQLite writes happen outside the step()
loop so they shouldn't affect GPU-parallelized throughput.

Tested on FetchPush-v4 (different env, same idea): ~42% to ~67% over
300 episodes with heuristic policies. Haven't benchmarked on ManiSkill
GPU envs yet — would need to check the overhead doesn't interfere with
the parallelization.

Anyone tried persisting episode-level metadata across training runs?

https://github.com/robotmem/robotmem