LIRO Stage 1 results

Snapshot published 2026-10-13 01:29 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.

No benchmark has finished yet. The queue on GPU 1 writes <run>/bench/summary.md per arm; this page picks them up when it is rebuilt.

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.