Add Unreleased changelog and document experimental features - #37
Merged
Conversation
Adds an Unreleased section covering the six PRs merged since v0.2.0 plus the RecGen work in #28, so the 0.3.0 notes are not reconstructed from git log later. Documents the two experimental features, which had no coverage at all, on a new Experimental Features page: place-next-to and RecGen shape completion. Both are off by default and share the same opt-in framing, and RecGen's setup, config, and caveats belong together rather than split across three pages. Server setup links out to the recgen repo instead of duplicating it. Cross-linked from the convex-hull section of Limitations, which describes the limitation RecGen mitigates. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Point readers at the experimental page from Limitations, Getting Started, and the Evaluation Workflow, so the features are discoverable from the pages where they are relevant rather than only from the nav. Add a feedback callout on the experimental page: these need reports from real setups to graduate, so link the issue tracker and the templates in Contributing. Drop the RecGen mention from Installation — it is a distraction at the point where someone is still getting the base system running. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The "10-20s per object" figure was carried over from the PR description and does not describe what the logs show. Reconstruction time depends on object complexity and GPU, and concurrency across GPUs dominates: total time tracks the slowest object, not the sum. Measured across the recorded runs on 4x 3090s: ~9s for a single object, 9-13s for 4-object scenes, 17-24s once requests start queueing at 5-6 objects. Documented with the hardware stated, and noted that more GPUs help more than faster ones. Also drop the "battle-tested" framing, note RecGen is a start rather than a fix, remove the issue-template pointer, and cut the em dashes. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
State the runtime in terms of the 4 GPUs running 4 requests in parallel, so it is clear why scenes with more than 4 objects take longer. Reword the Limitations pointer in plainer language. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Correct the near-placement description: the distance threshold is derived per object pair from their xy extents plus a fixed gap, not the 5cm value in default_constraint_to_tol, which is the constraint tolerance. Reference cuTAMP by file and symbol name rather than line numbers. Replace the invented known-issues list with the one point that matters, that the shipped q_capture is top-down and RecGen prefers a front-facing view, and give a concrete pose. Correct the multi-GPU claim: a scene reconstructs in one pass only when it has no more objects than the server has GPUs. Note RecGen takes a mask and intrinsics alongside RGB-D, and that the gateway defaults to port 18324. Use real place-next-to instructions from labelled successful runs, drop the 10-20s changelog figure, and reword the convex hull limitation, which described a distinction between inside and on top that the planner cannot express either way. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
#27 added support for goals that only pick an object: the prompt emits a holding atom, create_tamp_environment grounds it and omits HandEmpty, and the executor handles the resulting plan. 13 of our 15 labelled pick-only runs succeeded. Note that the robot ends the run holding the object and asks you to catch it before opening the gripper, since that is the part a user would not expect. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2 tasks
williamshen-nz
commented
Aug 3, 2026
Replaces the Unreleased heading and its compare link, which pointed at HEAD.
Vary the place-next-to examples and drop the claim about which synonyms are recognized, which is the LLM's behavior rather than ours. Document the 5cm satisfaction tolerance alongside the distance threshold, and link the cuTAMP sources on GitHub. Drop the open-box aside and the closing line about RecGen being slower. US spelling throughout.
Naming the two features here duplicates the page itself and dates the card as features come and go.
Revert the reworded convex hull example and the place-next-to aside, which sat under a stacking limitation it has no bearing on. Describe the RecGen mesh as predicted rather than reconstructed, since the point of that paragraph is what the sensor cannot see.
ryanlindeborg
approved these changes
Aug 3, 2026
ryanlindeborg
approved these changes
Aug 3, 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
3 participants
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.
Summary
Supersedes #36, which GitHub auto-closed when
will/recgenwas deleted on merge of #28. Same commits rebased ontomain— a closed PR's base cannot be retargeted, hence the new number. Review history is on #36.Adds an
[Unreleased]CHANGELOG section covering everything since v0.2.0, and a newdocs/experimental.mdfor the two experimental features (place-next-to, RecGen), neither of which had any docs. Cross-linked from Limitations, Getting Started, and Evaluation.The RecGen runtimes on that page are measured across the 26 recorded runs, not the "~10-20s per object" estimate from #28: requests fan out across GPUs, so total time tracks the slowest object until they start queueing.
Also moves pick-only instructions from the "bad" to the "good" list in Getting Started — #27 made them work.
Test plan
pixi run docs-build— clean; the 5 warnings are pre-existinglimitations.mdheading jumpsNotes
limitations.mdheading warnings are a one-character fix each, but would change existing anchor URLs, so left alone.