Skip to content

Release 0.7.1: armature, polynomial couplings, frame kinematics - #51

Merged
plancherb1 merged 52 commits into
mainfrom
codex/release-0.7.1
Nov 12, 2026
Merged

plancherb1 merged 52 commits into
mainfrom
codex/release-0.7.1

Conversation

@plancherb1

Copy link
Copy Markdown
Member

Release 0.7.1. One branch carrying the armature work and the wave that was planned as 0.8, recorded under one fresh receipt.

What is in it

  • Rotor armature: a baked armature= initializer with optional runtime tables on the native, NumPy, JAX and PyTorch surfaces, checked against Pinocchio's model.armature for inverse dynamics, the inertia matrix, its inverse, forward dynamics and the energies. A compact armature regressor and the forward-dynamics sensitivity to armature for identification. Supersedes Add per-joint armature (reflected rotor inertia) to fixed-base codegen #50; thanks to Felix Mühlenberend for the prototype and the TIAGo feedback.
  • Polynomial couplings (the MuJoCo joint equality), the first step of closed loops by constraint embedding, across the dynamics kernels and their gradients, the pose and frame Jacobian families, the integrators, the regressors, the centroidal quantities, Coriolis and the external-force gradients.
  • Frame velocity and acceleration with their derivatives on the three Python surfaces.
  • Integrator backward on floating-base and spherical models.
  • The first Jetson AGX Orin receipt and the split driver's merge of unsigned shards.

Output changes from 0.7.0

  • minv and the forward-dynamics torque cotangent on mimic and spherical robots carried doubled off-diagonal entries in Python on 0.7.0. They are fixed; a native consumer reading the raw buffer on such robots now sees zeros below the diagonal.
  • MuJoCo-convention integrators on a floating base take the MuJoCo step, so values and gradients differ.
  • The integrator retract normalises its input quaternion.
  • rizon4 inertials follow the vendor table.
  • Thread overrides clamp to the compiled kernel limit on every surface.

Known limit

On a free-floating arm at very high accelerations the float32 fdsva_so kernel returns two to three significant digits. The behaviour is the same as in 0.7.0 and the double build agrees with the reference. The diagnostic that compares floating fdsva_so with the reference stays opt-in and outside the receipt. The follow-up is an algorithm-level study in float32; no mixed precision is included.

Evidence

  • Receipt recorded at 29080a0, committed at 4339b69: 2264 tests passed, no skips, no carried shards, verified under test/gpu-proof-policy-release.yaml.
  • The tip adds one changelog commit outside the fingerprint; the receipt verifies there unchanged.
  • Submodules: RBDReference 6cda88c, URDFParser eb8ae03, GLASS 6a82320.

🤖 Generated with Claude Code

…lpha|; polynomial followers generate the baked fold
…Hessian twins follow; the surfaces no longer refuse the backward
…an; multitarget runners clamp to the launch bound
@plancherb1
plancherb1 merged commit 03cb47f into main Nov 12, 2026
12 checks passed
@plancherb1
plancherb1 deleted the codex/release-0.7.1 branch November 14, 2026 05:33
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.

1 participant