Skip to content

Add Halo loco-manipulation subset (55 takes, 5 behavior categories) - #3

Closed
bilal-bdai wants to merge 14 commits into
mainfrom
bilal/halo_locomanipulation_data
Closed

bilal-bdai wants to merge 14 commits into
mainfrom
bilal/halo_locomanipulation_data

Conversation

@bilal-bdai

Copy link
Copy Markdown
Collaborator

Source: RAI Halo Loco_Manipulation collections 10_27_2025 and 11_20_2025, converted to hold's BVH convention

Layout: bvh/halo_loco_manipulation// with five behavior buckets:

pick 33 originals + 33 mirrors
place 5 originals + 5 mirrors
carry 5 originals + 5 mirrors
tumble 5 originals + 5 mirrors
door 7 originals + 7 mirrors

Only the original skeleton size is provided (no 80%-scale variant in this subset). Each original ships with a _mirrored.bvh companion generated using the L/R-swap + axis-mirror convention used elsewhere in the dataset (validated byte-for-byte against existing hold mirror pairs).

Props parented to Z_UP_WORLD_ORIGIN with 8 corner end-sites per hold's prop-bone convention: Box, BoxSmall, BoxLarge, StepBox, Chair. Sole bones are estimated per-take with the velocity + height stance heuristic.

README updated with category table and per-category file index. Still missing the previews though

farbod-farshidian and others added 13 commits May 28, 2026 13:07
Source: RAI Halo Loco_Manipulation collections 10_27_2025 and 11_20_2025,
converted to hold's BVH convention by the cast pipeline
(cast.datasets.converters.halo_to_convention).

Layout: bvh/halo_loco_manipulation/<category>/ with five behavior buckets:

  pick    33 originals + 33 mirrors
  place    5 originals +  5 mirrors
  carry    5 originals +  5 mirrors
  tumble   5 originals +  5 mirrors
  door     7 originals +  7 mirrors

Only the original frame rate is provided (no 80%-scale variant in this
subset). Each original ships with a _mirrored.bvh companion generated using
the L/R-swap + axis-mirror convention used elsewhere in the dataset
(validated byte-for-byte against existing hold mirror pairs).

Props parented to Z_UP_WORLD_ORIGIN with 8 corner end-sites per hold's
prop-bone convention: Box, BoxSmall, BoxLarge, StepBox, Chair. Sole bones
are estimated per-take with the velocity + height stance heuristic.

README updated with category table and per-category file index.
The E-grab-box-from-shelf_000 take originally shipped without its
_boxsmall companion BVH, so the converted file in pick/ only carried
the StepBox (chair-as-step) prop. The boxsmall sidecar was recovered
upstream and the take has been re-run through the cast Halo converter,
which now merges both BoxSmall (the shelf box being grabbed) and
StepBox into the same human take.

Mirror regenerated to match.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

6 participants