Skip to content

Tolerance keys, external forces on every robot, osc_inertia backward error, g1 integrator bound, racecheck baseline for sm_87 - #42

Merged
plancherb1 merged 4 commits into
mainfrom
next-batch-2026-10-20
Oct 21, 2026
Merged

plancherb1 merged 4 commits into
mainfrom
next-batch-2026-10-20

Conversation

@plancherb1

Copy link
Copy Markdown
Member

Summary

  • The external-force CUDA modules run their pinocchio leg on every robot. gen3 was left out because the modules passed tolerance keys RBDReference does not have (forward_dynamics, f_ext_grad, f_ext_grad_so; the parameter-gradient module passed fd), which fell back to its 1e-7 default; the mimic robots were left out because the pinocchio oracle placed per-body wrenches through a mimic-free joint list. RBDReference PR Fix Python Bindings + Descriptions #15 (pin c5c23f8) closes the table, adds the forward_dynamics and inverse_dynamics_gradient buckets, maps a wrench on every body, and checks external forces on every robot against pinocchio. The GRiD modules now use the table's names.
  • test_osc_inertia checks every sample through the normwise backward error of the task-space inertia, the check the frame-Jacobian suite uses, instead of passing over ill-conditioned samples.
  • g1 floating joins the integrator gradient's float32 norm-bound class with a bound from the suite's noise model and the Jetson AGX Orin's measured class noise: at the most energetic RK4 sample the sm_87 gradient kernel rounds three times wider than the RTX 5090 at one stage state, and nothing else differs between the two GPUs. Guide entry 7.z55; the measurement scripts and logs are in the plan ledger.
  • compute-sanitizer racecheck is unsupported on the Orin's integrated GPU ("GPU debugging features are disabled"); the racecheck cell skips on that exact message and the sm_87 receipt row carries its own expected-skips baseline listing that one cell. The primary baseline stays empty.
  • Receipt refresh at the new pin: all 44 shards re-run fresh (the pin move stales every oracle-dependent shard), 1796 tests passed, verified under both the everyday and the release policy.

Verification

  • CPU lane 1253 passed. RBDReference fast suite 859 passed, slow external-force cells 72 passed.
  • RTX 5090: external-force three-way cells for gen3 (both bases), fr3 fixed and h1_2 floating; gen3 external-force gradients; racecheck; osc_inertia; g1 floating integrator, all pass. Full refresh receipt committed (86017f9).
  • Jetson AGX Orin at main: the g1 floating RK4 cell is the only deviation and sits at 0.58 of the new bound; every other sample, integrator and step matches the 5090.

🤖 Generated with Claude Code

@plancherb1
plancherb1 merged commit 0fb5923 into main Oct 21, 2026
12 checks passed
@plancherb1
plancherb1 deleted the next-batch-2026-10-20 branch October 21, 2026 15:29
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