NOTICE

EquiDexFlow
===========

EquiDexFlow derives from EquiGraspFlow (Lim et al.), used under the MIT License
(see LICENSE). The SE(3)-equivariant flow-matching backbone and several utility
modules originate upstream; the dexterous extension (joint contact/normal/force
prediction, differentiable hand kinematics, physics losses, the dexterous
trainer and scorer) is new in this repository.

Substantially upstream-derived (EquiGraspFlow):
  src/equidexflow/models/vn_layers.py
  src/equidexflow/models/vn_dgcnn.py
  src/equidexflow/models/vn_vector_fields.py
  src/equidexflow/models/equi_grasp_flow.py
  src/equidexflow/utils/Lie.py
  src/equidexflow/utils/ode_solvers.py
  src/equidexflow/utils/distributions.py
  src/equidexflow/loaders/acronym.py
  src/equidexflow/configs/equigraspflow_full.yml
  src/equidexflow/configs/equigraspflow_partial.yml

New in EquiDexFlow:
  src/equidexflow/models/equi_dex_flow.py
  src/equidexflow/models/{contact,force,hand_q,hand_q_flow}_decoder.py
  src/equidexflow/kinematics/  (differentiable Allegro/LEAP FK + physics terms)
  src/equidexflow/loaders/dexgrasp_db.py
  src/equidexflow/physics/scorer.py
  src/equidexflow/trainers/dex_trainer.py

The default training data is produced using a modified local fork of FRoGGeR (Li et al.), but EquiDexFlow does not depend on FRoGGeR at runtime: it consumes a documented JSON grasp
schema (see data/README.md) that any synthesis backbone can emit.

Allegro hand description (bundled in the package and at assets/hands/allegro/):
  src/equidexflow/_allegro_hand/  (== assets/hands/allegro/: allegro_rh.sdf +
  visual meshes) is vendored UNMODIFIED from FRoGGeR (Li et al., IROS 2023;
  https://github.com/alberthli/frogger), used under the FRoGGeR repository's MIT
  License (verified byte-identical, sha256 924dbf4d…, to alberthli/frogger
  main/models/allegro/allegro_rh.sdf). FRoGGeR in turn adapted it from Drake's
  allegro_hand_description (https://github.com/RobotLocomotion/drake, BSD-3-Clause)
  by renaming links/joints, as noted in the SDF header. We modified neither.

Object meshes redistributed via scripts/download_assets.py:
  - YCB object set (Calli et al., 2015) — https://www.ycbbenchmarks.com/
    Redistributed under CC BY 4.0. The specific .obj files we ship are the
    watertight "clean" variants produced by the FRoGGeR preprocessing
    pipeline (Li et al., IROS 2023; https://github.com/alberthli/frogger),
    used here under the FRoGGeR repository's MIT License. The underlying
    geometry is the YCB benchmark set; FRoGGeR's contribution is the
    cleaning/watertight remeshing step. Both upstreams must be credited
    when redistributing or building on this asset tarball.
  - EGAD! eval set (Morrison, Corke, Leitner, 2020) — https://dougsm.github.io/egad/
    Redistributed under CC BY-NC 4.0. Non-commercial use only.