E37: from the same clone, fpo-policy learns square under DPPO and falls under our FPO - #89
Merged
Merged
Conversation
…-ratio' and 'origin/feat/dppo-square-variant' into exp/e37-square-fpo-policy
…' into exp/e37-square-fpo-policy
…ur-cloned start E35's clone (0.44 / 0.38 success), observation statistics frozen (#82), the critic from the run's own initialisation; fpo-square under FPO with FPO++'s square fine-tuning (60 iterations of 48,000 steps), dppo-square under dppo square with E33's flow-policy changes (40 of 80,000); three seeds each. Learns: success over the last ten iterations at least +0.2 over the iterations the clone collected unchanged, on 2 of 3 (E34's rule). Written after the two pilots in results/pilot.txt, before the registered run.
…r DPPO (3 of 3) and falls under FPO (0 of 3)
This was referenced Sep 28, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
E37 was pre-registered at fb93a91, after two pilots. It fine-tunes E35's behaviour-cloned
fpo-policyon robomimic square under both algorithms: three seeds, three client processes each, onguangzhao.dppo-square(40 iterations)fpo-square(60 iterations, FPO++'s square settings as far as our FPO has them)fpo-policy· DPPO · square learns, 3 of 3.FINDINGS also records something read from the logs, not registered. FPO's critic never fits the return. Its value loss went only from about 9×10⁴ to about 4×10⁴ over 60 iterations, and the raw advantage std stayed near 400. DPPO's explained variance went from 0.40 to 0.61. The FPO arm gave its critic the actor's learning rate (1e-5, with rewards ×10). FPO++ gives its critic 1e-4. PROTOCOL.md declared that as a known deviation, and it is the first thing to try next. By this project's rule this is a defect of ours, not a finding about FPO.
The figure update (
fpo-dppo-square→ learns,fpo-fpo-squarefrom E37's runs) will go into #88 and the site PR.