feat(trim): add motion trimming - find dead frames from proprioception - #39
Merged
Conversation
Two modules. proprio/ turns the robot's joint-state columns into a per-frame motion signal: motion_scale measures the typical per-dim step in one pass, motion_energy diffs each dim against its previous value inside the episode and combines them as a normalized Euclidean norm, is_active thresholds that with a run requirement so one noisy frame is not motion. All plain Daft expressions - no UDF, no video decode. trim/ reduces that to one contiguous window per episode. It takes a DataFrame and returns a DataFrame, unlike the rest of the package, because a window is an aggregation across an episode's rows rather than a per-row value. Aborted takes where the arm never moves keep their full span and get a never_active flag. The two views serve different consumers: frame samplers filter on is_active and drop interior pauses too; video-slice consumers take the window, which has to keep them to stay contiguous. Validated on DROID 1.0.1 (4 shards, 4,391 episodes, 1.28M frames): 3.8% trimmed in the window view, 11.2% per-frame, matching the multibase prototype. 25 new tests, 103 passing overall.
daft-physical-ai trim scaffolds a motion-trimming demo (script + notebook + markdown) from a TrimDemoConfig, mirroring hands and rewards. The committed examples/trim/ is the executed default: one DROID shard streamed from Hugging Face, 321,344 frames, 4.0% dropped in the window view and 11.2% in the per-frame view. 12 new tests.
""[:1] in "[{" is True (empty string is a substring), so an empty cell
rendered as a stray backtick pair in the markdown table.
frames.show(5) after the build cell, so all three formats display what one-row-per-frame actually looks like before the motion scoring.
The task column showed blank for episode 0 (~21% of DROID episodes have no task label) and the trim workflow never uses it. Also drop the score_rewards aside from the what-it-saves note.
ykdojo
marked this pull request as ready for review
July 30, 2026 19:39
ty infers Expression == int as bool (daft doesn't type __eq__), so the one live == comparison fails the style check; is_in is typed properly and already used elsewhere in the package.
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.
Finds the dead frames in an episode - operator setup before anything happens, the tail after the task is done - without decoding video. The robot's joint positions sit in parquet next to the mp4, so a still arm is a columnar scan away. On DROID that's 12 GB of proprioception against 400 GB of video.
Two outputs, for two kinds of consumer: a per-frame
is_activeflag (frame samplers - drops interior pauses too) and one contiguous window per episode fromtrim_windows(video slices).proprio/is plain expressions, same shape astrack_hands/score_rewards;trim/takes a DataFrame and returns one.On DROID 1.0.1 (4 shards, 4,391 episodes, 1.28M frames): 3.8% trimmed in the window view, 11.2% per-frame, reproducing the multibase prototype. Aborted takes where the arm never moves keep their full span, flagged
never_active.examples/trim/is the committed executed walkthrough (one DROID shard streamed from Hugging Face), generated by the newdaft-physical-ai trimCLI subcommand. 25 new tests, 103 passing.Prototype, method writeup, and figures: Eventual-Inc/multibase#1169 (private).