LIRO Stage 1 results

Snapshot published 2026-10-13 10:52 UTC; the campaign is still running, so numbers and videos of later snapshots differ. Campaign directory generated/liro_stage1_20261011, first run started 2026-10-11 20:04 UTC, code aa1e901e0, seed 73, 2273402 parameters in the base network. All ten arms train on this machine (ankac-dgxc-4gpu); the OSMO wave ordered on 2026-10-11 has not started because the dispatcher stopped reporting at 20:15 UTC that day.

1. The ten arms

Every arm is the base configuration with one change. The selection metric is the mean final residual of the full-horizon validation (4 held-out worlds, 8 queries per step, the horizon of the growth stage); lower is better and the trainer keeps the checkpoint of the best eligible epoch. "rho after 8" is the cheap validation's pose error ratio against the float64 reference after 8 queries on one step (1 = the predictor pose, mean over 8 worlds). Momentum drift is relative to the gravity impulse over the rollout. Resets count training worlds that tripped a guard and were replaced; the figure in brackets splits them by guard. Wall-clock is the sum of the epoch times; three or four trainers shared one GPU.

2. Training curves

Dashed lines are runs still training. The grey bands and the labels above each panel are the growth stages of the regime (the largest iteration count K and horizon H a training job can have); the validation horizon grows with the stage, so a metric is comparable with another only inside one band.

3. Benchmark

Held-out sets, best checkpoint of each arm: tier A is the error ratio against the float64 reference after K queries, the SolverVBD rows are the same ratio after n iterations of the Newton solver (compliant ALM and legacy variants; n = 512 shows the formulation floor), tier B is the rollout survival and energy. The benchmark runs on GPU 1 with no training on it, one arm after another.

Overview of all arms

One-step columns: the mean pose error against the float64 minimiser after 8 queries divided by the predictor's mean error over the set (1 = the predictor pose; this ratio of means is free of the floor effect that inflates the per-world geometric mean when the predictor is already within 1e-4 L_ref of the minimiser), and the median per-world ratio. Rollout columns: worlds surviving 300 steps at 8 queries, the median momentum drift of the free articulations and the median final residual.

SolverVBD compliant ALM on the same uniform-stiffness worlds (from the first benchmark that ran it): mean error ratio after 8 iterations 0.4629 (median per world 0.3441), after 32 iterations 0.3559 (median 0.1534).

Reading of the base arm's one-step error, split into parts (local checkpoint, 23 training-distribution worlds, 8 queries): the translation error falls from 0.00177 to 0.00120 L_ref (0.68 times) while the rotation error rises from 0.00129 to 0.00188 rad (1.45 times); after the queries every world sits at a rotation error of 0.0005 to 0.005 rad whatever its starting error, so the network's output has a rotation floor near 1e-3 rad. Rotations about the hinge axes are the directions the step energy constrains least.

base

local checkpoint, benchmark on GPU 1 (shared with another agent's training job, so the timing columns are inflated)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/base_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Equivariance check (largest pose discrepancy under a world rotation): 7.63e-06

cluster checkpoint, benchmark on the CPU (the timing columns are CPU times)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/cluster/base_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Equivariance check (largest pose discrepancy under a world rotation): 3.81e-06

learned_tr

local checkpoint, benchmark on GPU 1 (shared with another agent's training job, so the timing columns are inflated)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/learned_tr_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Equivariance check (largest pose discrepancy under a world rotation): 3.81e-06

cluster checkpoint, benchmark on the CPU (the timing columns are CPU times)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/cluster/learned_tr_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Equivariance check (largest pose discrepancy under a world rotation): 3.81e-06

no_tr

local checkpoint, benchmark on GPU 1 (shared with another agent's training job, so the timing columns are inflated)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/no_tr_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Equivariance check (largest pose discrepancy under a world rotation): 3.81e-06

cluster checkpoint, benchmark on the CPU (the timing columns are CPU times)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/cluster/no_tr_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Equivariance check (largest pose discrepancy under a world rotation): 1.91e-06

art_frame

local checkpoint, benchmark on GPU 1 (shared with another agent's training job, so the timing columns are inflated)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/art_frame_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Equivariance check (largest pose discrepancy under a world rotation): 7.63e-06

cluster checkpoint, benchmark on the CPU (the timing columns are CPU times)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/cluster/art_frame_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Equivariance check (largest pose discrepancy under a world rotation): 3.81e-06

all_pair

local checkpoint, benchmark on GPU 1 (shared with another agent's training job, so the timing columns are inflated)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/all_pair_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Equivariance check (largest pose discrepancy under a world rotation): 7.63e-06

cluster checkpoint, benchmark on the CPU (the timing columns are CPU times)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/cluster/all_pair_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Equivariance check (largest pose discrepancy under a world rotation): 5.72e-06

gn_input

local checkpoint, benchmark on GPU 1 (shared with another agent's training job, so the timing columns are inflated)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/gn_input_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Equivariance check (largest pose discrepancy under a world rotation): 1.86e-05

cluster checkpoint, benchmark on the CPU (the timing columns are CPU times)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/cluster/gn_input_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Equivariance check (largest pose discrepancy under a world rotation): 1.84e-05

gn_anchor

local checkpoint, benchmark on GPU 1 (shared with another agent's training job, so the timing columns are inflated)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/gn_anchor_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Equivariance check (largest pose discrepancy under a world rotation): 7.63e-06

cluster checkpoint, benchmark on the CPU (the timing columns are CPU times)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/cluster/gn_anchor_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Equivariance check (largest pose discrepancy under a world rotation): 8.46e-06

teacher

local checkpoint, benchmark on GPU 1 (shared with another agent's training job, so the timing columns are inflated)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/teacher_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Equivariance check (largest pose discrepancy under a world rotation): 7.63e-06

cluster checkpoint, benchmark on the CPU (the timing columns are CPU times)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/cluster/teacher_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Equivariance check (largest pose discrepancy under a world rotation): 5.72e-06

ik_fk

local checkpoint, benchmark on GPU 1 (shared with another agent's training job, so the timing columns are inflated)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/ik_fk_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Sets not scored:

- four-bar loop: the reduced arms have no closed-loop route (spec 3)

Equivariance check (largest pose discrepancy under a world rotation): 7.63e-06

cluster checkpoint, benchmark on the CPU (the timing columns are CPU times)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/cluster/ik_fk_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Sets not scored:

- four-bar loop: the reduced arms have no closed-loop route (spec 3)

Equivariance check (largest pose discrepancy under a world rotation): 9.54e-06

reduced_learned

local checkpoint, benchmark on GPU 1 (shared with another agent's training job, so the timing columns are inflated)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/reduced_learned_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Sets not scored:

- four-bar loop: the reduced arms have no closed-loop route (spec 3)

Equivariance check (largest pose discrepancy under a world rotation): 5.72e-06

cluster checkpoint, benchmark on the CPU (the timing columns are CPU times)

LIRO benchmark summary

checkpoint: generated/liro_stage1_20261011/cluster/reduced_learned_s73/checkpoints/best_validation.pt

Tier A: error ratio after K queries (1 = the predictor pose)

Geometric mean over the worlds whose float64 reference converged. Every ratio is clipped to [0.001, 10]; a non-finite ratio becomes 10 and counts in the failed column.

SolverVBD on the uniform-stiffness set: error ratio after n iterations

The n=512 column is the formulation floor, not a failure to converge: LIRO's hinge angular row uses the deviation form of kappa and SolverVBD uses the plain logarithm, so the two energies have different minimisers. That floor measures about rho 0.09 on the uniform-stiffness set.

Tier B: rollouts of 300 steps at K = 8

Survival is the trainer's rule: finite energy below the blow-up factor times the loss floor, and every anchor gap below gap_fail L_ref. The energy column is the geometric mean over the surviving worlds.

Sets not scored:

- four-bar loop: the reduced arms have no closed-loop route (spec 3)

Equivariance check (largest pose discrepancy under a world rotation): 5.72e-06

4. Videos

Left: the network with K queries per physical step. Middle: SolverVBD (compliant ALM) at the same number of iterations. Right: SolverVBD at many iterations, the converged reference. All three start from the same state and see the same prescribed base motion. One video frame per physical step, so the playback is slow motion.

5. Anchor gap over the rollouts, network against SolverVBD

From the rollouts of section 4: the largest joint anchor gap of the world after every physical step, measured with the same joint geometry for every solver (the distance between the two anchor points of a joint, in units of the reference length L_ref; the trainer fails a world above 3 L_ref). An anchored chain under load has a physical gap, the stretch of its soft joints, so the question is whether the network reproduces the gap of the converged solver, not whether it reaches zero.

6. Cluster copies

The same ten configurations and seed ran on the OSMO cluster, one L40 each (dispatcher inbox entry 26, campaign liro1, code aa1e901e0). They are a second sample of every arm: the same code and data, a different GPU and a different order of floating-point sums, so the curves differ from the local runs where training is sensitive. Reports are pulled from the branches osmo/liro/liro1/<arm>_s73.

7. Notes

• Normalised units: lengths in the reference length L_ref, masses in the reference mass m_ref of the world, dt = 1. The residual is the RMS over an articulation's dynamic bodies of the six gradient entries of the step energy.
• The first three arms (base, learned_tr, no_tr) trained together on GPU 3 from 2026-10-11 20:04 UTC and finished in about 3.5 hours. The other seven started 2026-10-12 22:24 UTC: four on GPU 3, three on GPU 2.
• Source of every number: <run>/report.json (per-epoch records) and <run>/bench/summary.md. Rebuild this page with python -m experiments.liro.results_page.