Persisting training insights across MuJoCo runs

[robotmem]

Hi,

I've been training manipulation policies with MuJoCo (FetchPush, custom
envs with Franka) and noticed a recurring pattern: after tuning reward
weights or contact parameters and restarting training, there's no easy
way to keep track of what configurations worked in previous runs.

Not talking about checkpoints or replay buffers — more like high-level
notes: "kp=200 caused oscillation", "approach angle of 15° was stable",
that kind of thing. Currently I just keep a text file, which doesn't
scale.

I put together a small Python wrapper that logs episode summaries
(reward, steps, success, key params) to a local SQLite db and lets you
query past experience at the start of a new run:

from robotmem.gymnasium import RobotMemWrapper

env = gym.make("FetchPush-v4")
env = RobotMemWrapper(env, collection="fetch_push")
# episode outcomes auto-logged on reset
# past tips available via info["robotmem_tips"]

On FetchPush-v4 with heuristic policies (10 seeds), keeping a running
log of past episode outcomes improved success rate from ~42% to ~67%
over 300 episodes. Effect varies a lot by task complexity though.

Has anyone tried something similar for tracking experiment history in
MuJoCo? Curious if there's a standard approach I'm missing.

https://github.com/robotmem/robotmem