ORA Appendix

Protocol

What changed in the full rerun

On the 399 cases with valid outcomes in both LIBERO tables, the refreshed run rescued 36 failures and lost 34 previous successes (exact McNemar p = 0.91). On 796 LIBERO-Pro pairs, it rescued 67 and lost 52 (p = 0.20). The small overall increases do not establish a reliable gain. This rerun includes several harness changes and cannot isolate the effect of overhead localization. It replaces the earlier, incomplete progress readout.

Study identifiers

Sources

RPent percentages and available denominators, including contributor corrections and the updated GPT-5.5 Long results, come from the pinned leaderboard page and its versioned result data (revision 705bd44b, checked 2026-09-22). ORA counts come from queue records, with each counted run's plan, result, evaluation and seal hashes checked; snapshot 2026-10-12 11:25 PDT. Data and export scripts: page source.

Exact values by benchmark and suite

RoboTwin C2R

LIBERO-Pro

LIBERO

Ours: rate over all cases, a case without a valid outcome counting as a failure; suite cells are successes / valid outcomes. — not reported; missing suites are not inferred. [a] Astra uses separately frozen exploration-memory snapshots for Long (157/200) and the other six suites (584/600). [b] Flash Mode uses released GPT-5.5 exploration memory and reports the best-performing seed from s0–s9. [c] ASPIRE's 61.36% is the board's mean of eight reported percentages; its Long results use zero-shot transfer from a LIBERO-90 skill library^[1]. The suite table also keeps RPent's motor-only Astra result (38% on Long Task only), which uses neither VLA primitives nor Memory.

	LIBERO	Pro · Task	Pro · Swap
Spatial	95/100	98/100	91/100
Object	92/100	93/100	98/100
Goal	84/100	77/100	78/100
Long	63/100	52/100	64/100

Successes / valid official outcomes. Long is the weakest family in all three settings.

Cost

Historical cost samples cover the earlier g003/g004 benchmark cohorts and the RoboDojo scan. They do not estimate the refreshed g005 cohort. Token usage is priced at CNY 0.02 / 1.00 / 4.00 per million cached input, uncached input and output tokens, with 7.1 CNY per USD (recorded September 17 rates). The self-hosted service does not report cache hits. The cost script infers cacheable tokens from request lengths, so the estimated 84% to 86% cached-input share is an assumption, not a measured hit rate. Simulator and GPU costs are excluded.

Earlier API runs in September: $0.38 per randomized RoboTwin episode at off-peak list prices ($0.76 at peak, 44 episodes), $0.29 per clean episode (391) and $0.37 per LIBERO-Pro episode (164).

Failure audit and time

Figure A1. Main cause of each failed RoboDojo round. EXP11C scan, source c6b08a06, one auditor judgement per round; 27 of the 64 motion cases are low-confidence. No round exhausted its decision budget. Most failed rounds ended when the model stopped immediately after a failed motion. The auditor judged recovery still possible in only three rounds. In two earlier scans, 91% of 746 failed inverse-kinematics attempts found no solution from 32 seeds although the targets were inside the workspace bounds.

Figure A2. Where episode time goes. Time shares come from EXP12B g002: 79 episodes, 12,206 model calls and 91.7 episode-hours, using paired trace timestamps. The episode medians come from different, later cohorts (RoboTwin g004; LIBERO and LIBERO-Pro g005): RoboTwin 25.0 min, LIBERO-Pro 15.9 min, LIBERO 13.8 min. A paragraph-level classifier labelled 58% of reasoning paragraphs as concerning geometry, a broader category and a different unit from responses that compute geometry by hand. The audit estimated that batching independent read-only calls could remove at most 7% of model calls in this sample.

Media and case provenance

Clips and the replay use original captured frames, not simulator reconstructions, at simulation time with model and tool waiting removed. Text panels show visible model text, tool calls, results and receipts only. Hero tiles run from reset to official success at native resolution or above: RoboTwin tiles are clean scenes from EXP12B warm-up sessions (not evaluation runs), LIBERO and LIBERO-Pro from g003, RoboDojo from the first scans, frame-interpolated from 10 Hz. Manifests and hashes: page/assets/fresh-v41/media-manifest.json, teaser/v12/panels.json.

Both are two-round sessions on a repeated seed (source c58d6b0), task memory off, chosen after inspecting outcomes rather than sampled. The laptop quotes are the model's visible text, events 59, 348, 396 and 426; receipt hashes are in page/cases-v41.json.