Repository navigation
Home page: the cost with reuse-feedback-obs, other trainers, the environment side against RLlib and LeRobot, a quickstart that runs - #25
Merged
Conversation
…gainst RLlib and LeRobot, a quickstart that runs - What the split costs: each observation now crosses once (reuse-feedback-obs, protocol version 2); E46's numbers replace E43's "twice the observation's bytes". - New section: RLinf, Stable-Baselines3 and CleanRL train PlugRL env clients (E48, E49), with E49's curves from the bridge's episode logs; SB3 ends on byte-identical weights behind the protocol (E50, E52, E53). - The environment side is light: E45's comparison with RLlib's external env and LeRobot's actor, and what a thin environment side alone does not show. - Quickstart: `uv run` inside each repository, as Get Started has it; the bare `plugrl-run-server` was not on PATH after `uv sync`. English and Chinese pages alike.
…sion 1, not a version 2 SPEC.md section 10: the protocol is version 1, and reuse-feedback-obs is a feature the server lists and the client uses. E46's "v1" and "v2" are its arms, not protocol versions.
tactino
changed the title
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
The home page had stood still since 09-29. This brings it to what is merged and measured now. English and Chinese pages alike.
Wrong before, corrected
reuse-feedback-obs, a negotiated feature of protocol version 1 (SPEC.md section 10.1) that plugrl-server offers and plugrl-env-client uses by default. The cost section now uses E46: a 184 KiB camera observation costs 11.6 ms per step on a 23.7 MB/s link, against 18.5 ms without the feature, and the weights are byte-identical either way.plugrl-run-serverafteruv sync, which leaves it inside the repository's.venv, so the command was not found. It now usesuv runinside each repository, as Get Started does.New
plugrl-bridges, 3 of 3 seeds in every pairing (E49), and RLinf trains env clients on a laptop with no torch, Ray or RLinf (E48). A figure shows every seed's curve (docs/media/other-trainers.png, drawn byexperiments/e49-multi-trainer/figure.pyin plugrl-server from the bridge's episode logs). The page says plainly that RLinf learns Pendulum later than SB3.plugrl-env-clientis 222 MB with no torch; RLlib's external-env client and LeRobot's actor are about 6.1 GB each. The section also says what this does not show: dm_env_rpc and openpi are smaller, but neither trains through its channel.plugrl-bridgesis not public yet, so the page names it without a link and links the experiments instead.Check: the site's CI (
mkdocs build --strict) on this PR.