docs: show the VLA, do not only describe it - #9
Merged
Merged
Conversation
An eval-only capture on the cluster: one env client on libero_spatial task 0, three episodes, all three successful. The recorder writes the observation stream the client sends to the server, so the clip is what the policy sees at its native 224x224, not an outside camera. It plays under the VLA heading in both languages, captioned as the unmodified checkpoint - a robot succeeding directly above a section headed 'the reinforcement learning result is negative' would otherwise read as the trained policy, which it is not. Paths are root-absolute: the build puts media at site/media only, so a relative path would break the Chinese page under /zh/. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 said a full-size pi0.5 runs through the boundary. Now it shows it.
What was captured
An evaluation run on the cluster - the
evalalgorithm moves no weights - with one env client process onlibero_spatialtask 0, three episodes, all three successful. The env client's recorder was switched on (--recorder.episode-freq 1 --recorder.record-video --recorder.record-full-rollout), which writes the observation stream it sends to the server: one mp4 per camera, at the policy's native 224x224.So the video is not an outside camera filming a simulator. It is what the policy saw, arriving over the same WebSocket connection the actions came back on. 301 env steps, 61 inference calls, 61 feedback calls.
What is on the page
The base camera plays under the "A real VLA, end to end" heading, in both languages, with a caption that says plainly:
That sentence is the point of the caption. A reader who sees a robot succeeding directly above a section headed "the reinforcement learning result is negative" will otherwise assume the clip shows the trained policy. It does not; it is the baseline.
Paths
/media/..., root-absolute, not relative. The Chinese page is served from/zh/, andmkdocs buildputs the media atsite/mediaonly - there is nosite/zh/media- so a relative path renders a broken video on exactly one of the two languages. Verified against the build output rather than assumed.Also in
docs/medialibero-wrist.mp4, the second camera, andlibero-demo.gif, a 934 KB loop of the same episode. Neither is used by the site yet. The GIF is here so theplugrl-serverREADME can point athttps://plugrl.github.io/media/libero-demo.gifinstead of carrying a binary of its own - that repository's pre-commit hooks cap added files at 500 KB, and shrinking the clip to fit would cost resolution for no reason.🤖 Generated with Claude Code