G1 FACTORY: FINAL PAIRED ENGINEERING EVALUATION 270 audited episodes; 90 matched task/reset conditions; 9 tasks x 10 seeds x 3 modes. This report compares the selected ACT ensemble, published GR00T checkpoint, and adaptive router. OVERALL NATIVE SUCCESS ACT ensemble 32/90 (35.6%) 95% Wilson reference interval [26.4%, 45.8%] GR00T 69/90 (76.7%) 95% Wilson reference interval [66.9%, 84.2%] Adaptive 70/90 (77.8%) 95% Wilson reference interval [68.2%, 85.1%] These are descriptive results for these checkpoints and fixed native scenes. Ten seeds per task are a small sample. Reference intervals do not cover scene, domain, training, or infrastructure uncertainty; the 90-condition aggregate spans heterogeneous tasks. PER-TASK NATIVE SUCCESS (k/n) Task ACT ensemble GR00T Adaptive LMBoxLift 6/10 10/10 10/10 LMBoxLiftFloor 3/10 10/10 9/10 LMBoxTableToShelfPnP 2/10 6/10 9/10 LMDrillLift 1/10 7/10 8/10 LMDrillLiftObs 2/10 9/10 9/10 LMDrillPnP 8/10 8/10 7/10 LMPickDrillFromHolder 6/10 7/10 7/10 LMPushButton 3/10 6/10 6/10 LMPushShelfForward 1/10 6/10 5/10 PAIRWISE MATCHED OUTCOMES A vs B: both succeed | A only | B only | both fail | (B-A) percentage points ACT ensemble vs GR00T: 25 | 7 | 44 | 14 | +41.1 pp; N=90 ACT ensemble vs Adaptive: 24 | 8 | 46 | 12 | +42.2 pp; N=90 GR00T vs Adaptive: 63 | 6 | 7 | 14 | +1.1 pp; N=90 MODEL SWITCHING AND CLOUD USE Mode Cloud queries / decisions Actual switches ACT / GR00T executed steps ACT ensemble 0/1778 (0.0%) 0 70490 / 0 GR00T 1278/1278 (100.0%) 0 0 / 49941 Adaptive 1122/1264 (88.8%) 84 9156 / 40074 Per-task cloud queries / decisions; actual switches: LMBoxLift: ACT ensemble 0/141, 0 switches; GR00T 72/72, 0 switches; Adaptive 67/72, 12 switches LMBoxLiftFloor: ACT ensemble 0/226, 0 switches; GR00T 114/114, 0 switches; Adaptive 99/125, 12 switches LMBoxTableToShelfPnP: ACT ensemble 0/227, 0 switches; GR00T 186/186, 0 switches; Adaptive 148/155, 9 switches LMDrillLift: ACT ensemble 0/234, 0 switches; GR00T 134/134, 0 switches; Adaptive 103/116, 12 switches LMDrillLiftObs: ACT ensemble 0/219, 0 switches; GR00T 135/135, 0 switches; Adaptive 133/136, 8 switches LMDrillPnP: ACT ensemble 0/141, 0 switches; GR00T 147/147, 0 switches; Adaptive 152/160, 11 switches LMPickDrillFromHolder: ACT ensemble 0/156, 0 switches; GR00T 137/137, 0 switches; Adaptive 127/138, 6 switches LMPushButton: ACT ensemble 0/191, 0 switches; GR00T 137/137, 0 switches; Adaptive 76/133, 8 switches LMPushShelfForward: ACT ensemble 0/243, 0 switches; GR00T 216/216, 0 switches; Adaptive 217/229, 6 switches A decision issues a 40-step action chunk at 50 Hz; the last chunk may terminate early. Actual switches count executed transitions between selected policies, excluding initial selection. Five-step cosine crossfades are logged. Query totals vary with episode length and are not fixed-time compute savings or network-bandwidth measurements. ACTION DEVIATION MEASUREMENTS Left wrist position RMS (mm) mean 23.9243; maximum chunk RMS 126.541; 1122 queried chunks Right wrist position RMS (mm) mean 33.1848; maximum chunk RMS 177.265; 1122 queried chunks Left wrist orientation RMS (deg) mean 5.71769; maximum chunk RMS 40.5205; 1122 queried chunks Right wrist orientation RMS (deg) mean 6.95011; maximum chunk RMS 41.1851; 1122 queried chunks Left hand RMS (rad) mean 0.0184353; maximum chunk RMS 0.710069; 1122 queried chunks Right hand RMS (rad) mean 0.0550943; maximum chunk RMS 0.728335; 1122 queried chunks Base-height RMS (m) mean 0.00532684; maximum chunk RMS 0.113156; 1122 queried chunks Navigation XY RMS (m/s) mean 0.0674066; maximum chunk RMS 0.420344; 1122 queried chunks Navigation yaw RMS (rad/s) mean 0.101599; maximum chunk RMS 0.845938; 1122 queried chunks These compare ACT and GR00T predicted commands on the same queried observation. They are not measured tracking error or object pose error. Wrist values use the validated Factory USD forward kinematics. Means weight queried chunks equally; unqueried disagreement is unknown. Full per-query series, issued commands, selected model, and transitions remain in the run evidence and visual simulation viewer. NATIVE TERMINATION AND INTERPRETATION ACT ensemble: success=32, time_out=58 GR00T: success=69, time_out=21 Adaptive: success=70, time_out=20 Native term counts can overlap when multiple termination predicates fire together. LMBoxLift: Object root world Z >0.18m and at least one finger link within0.2m. Absolute world root height plus finger proximity; no secure-contact or dwell requirement. LMBoxLiftFloor: Object root world Z >−0.5m, local-Z/world-up dot ≥0.9, and at least one finger within0.6m. Absolute root height and orientation; the0.6m finger proximity allowance is a broad proxy. LMBoxTableToShelfPnP: Object/destination root separation per axis <[0.16,0.14,0.9227]m, object static, every finger >0.1m away, and destination contact >0.95N. World-axis root proximity and destination contact; static permits speed up to0.15m/s and0.15rad/s. LMDrillLift: Drill root world Z >0.16m and approximate grasp predicate. Grasp is approximated by finger proximity and joint closure, without required object-hand contact. LMDrillLiftObs: Drill root world Z >0.16m and approximate grasp predicate. Success does not explicitly require obstacle avoidance. Approximate grasp; success does not explicitly test that the robot avoided colliding with the obstacle. LMDrillPnP: Drill/destination-table root separation <[0.3,0.27,0.81]m, static, local-Z/world-up dot ≥0.8, and every finger >0.1m away. No contact term. Root proximity, orientation and static test; no destination contact term. LMPickDrillFromHolder: Drill root world Z >0.13m, approximate grasp, and holder contact is not above0.95N. Approximate hand grasp plus no holder contact; removing the drill is not a secure-grasp guarantee. LMPushButton: Button joint normalized pressedness >0.5 and at least one finger within0.2m. Normalized joint travel plus finger proximity; no force-contact or sustained-press requirement. LMPushShelfForward: Shelf/target root separation <[0.3,0.3,0.01]m, static, local-Z/world-up dot ≥0.9, and a finger within1.1m. Root/target proximity and static/upright test; the1.1m finger proximity allowance is broad. Success retains the official geometric/proximity predicates; it does not certify secure grasping, collision avoidance, human-safe handover, or general humanoid reliability. No predicate was strengthened or loosened to favor a policy. TRAINING AND SELECTION LIMITS ACT consists of four independently trained full models with ResNet18 ImageNet initialization. Each received 40,000 optimizer steps at batch 16: 160,000 member-steps and 2,560,000 sampled examples in total. There were 90 training and 10 reserved episodes per task (810/90 overall); each member used 72 training episodes per task, 648 total, with equal-task sampling. ACT received a nine-way one-hot known task instruction and deterministic 320x240 RGB. GR00T is the published G1-Factory-GR00T-1.7 expert, already trained on all 900 public Factory episodes after foundation-model pretraining. No new GR00T fine-tuning was performed here. It receives native 640x480 RGB and the native language instruction. Pretraining, task-data exposure, optimization budget, instruction encoding, and preprocessing are unequal. Fixed-prefix equal-task normalized command MAE used for ACT checkpoint selection: 10,000 steps: 0.1424042215 20,000 steps: 0.1264782756 30,000 steps: 0.1213080474 40,000 steps: 0.1188063193 (selected) The 40k snapshot was selected before closed-loop outcomes, from the predeclared 10k/20k/30k/40k candidates. Prediction error was still improving at 40k; convergence or a fully optimized ACT baseline is not established. The same 90 reserved episodes and 900 prefixes were reused for selection and task-specific q95 dispersion calibration. This is not an untouched test split, and these demonstrations were not unseen by GR00T. The router is an empirical deviation-based engineering mechanism. Ensemble dispersion is not a calibrated probability of physical failure, and exact preservation of a target action distribution is not claimed. Results cannot establish architecture-only GR00T superiority. RUNTIME, INVALIDATION, AND RETRIES Inference used synchronous loopback RPC while simulation waited. No WAN latency, asynchronous latency hiding, Jetson/Orin execution, hardware handover, or real-time deployment was measured. Native 50 Hz simulation time does not imply 50 Hz wall-clock performance. The first 54 completed episodes (all nine tasks, seeds 1000 and 1001, all three modes) ran with one outer simulator worker. All 54 outcomes were retained. After an excluded seed 42 two-worker standing diagnostic passed, scheduling resumed with at most two outer workers for the remaining 216 episodes. Actual overlap varies with task duration and queue completion; an inner run receipt stating one worker does not describe global scheduling concurrency. The scheduling transition changed no model, controller, per-episode protocol, frozen runtime, or router setting. The diagnostic included physical-state/FK/clock checks and sampled RGB inspection, but did not retest learned-model RPC interleaving or prove universal pixel validity. These scheduling conditions are not a controlled concurrency-performance experiment. Wall-clock measurements span serial and partially overlapping execution as well as the approved logging-only runtime revision. Even within one runtime SHA, pooled timings cannot be interpreted as isolated hardware latency, network latency, real-time capability, or a demonstrated twofold speedup. Excluded v1 attempts: 21; all v1 results, including successes, are outside the primary cohort. V1 was invalidated after actual saved policy inputs revealed black reset cameras. V2 added pre-action visibility/readiness and physical-state guards. Later darkness/occlusion remains diagnostic and never removes a policy failure by itself. Approved v2 retained episodes: 24; v3 episodes: 246. Only the 24 immutable completed v2 episodes from the first eight seed 1000 task triples were retained. The approved v2-to-v3 proof removes exactly one scheduled traceback watchdog call from the simulator runner; model, controller, camera, task, and router bytes remain unchanged. Each retained run is audited against its original freeze. The watchdog was implicated by crash timing, but causation is not claimed proven. All three original LMPushShelfForward/seed1000 modes were superseded, including the two completed outcomes; the complete triple was rerun under v3. Their original artifacts remain archived. V2 and v3 wall-time latency is not pooled as one identical runtime. First fully audited completed result regardless of task outcome. At most two v3 infrastructure retries per case; further infrastructure failures stop for diagnosis. Task failures are never rerun. Final-root attempt status counts: {"validated_completed": 270} Infrastructure failures and incomplete attempts are reported separately from native task failures. The selected result is the first fully audited completed attempt by numeric index, irrespective of success. Seed 42 integration pilots are excluded. Adaptive episodes that executed only GR00T but differed from the GR00T baseline outcome: 6. Such differences cannot be attributed to switching. Client RNG isolation is verified, but the full rendering/physics system is not claimed bitwise deterministic. PROVENANCE Expert identity: {"backbone_manifest_sha256": "6c4f04a5cc74e2413b8648812cb1667994f259af8789c26c03b0a848bee4974e", "manifest_sha256": "c98f5211f0d10ba4cdcb5019e15d20949e37b4d03837d2826960007a55d1bd3b", "repo": "nvidia/G1-Factory-GR00T-1.7", "revision": "e0f5a056257fc89b386b7584cd5978285c6a02de"} Dataset: nvidia/G1-Factory-Benchmark @ 1c5b2fbfaa35f70451bb462426c1b8adec1b7ec6 Factory source: 81cf7909b13fe1a961aae9a4e457a5e786e4df8c Runtime freeze: 1b63b1d5495fd24fa8d9b558c81e4be971f3d7154d6ec769155bee5fa79a4658 Approved lineage registry: 9a644bb75428a0b72cd0e8e9d16770de8c7172dc8640aab9be3b24d9a1281cc5 Scheduling transition: f0552158d4079fc30a38e609138921f26d077aab3038c1913471e2e8336fc2c0 Summary: 83da13f3cb002ba30a6bff50e485de9d4659140e26375450e2534fccb7d852d6 Protocol: e08c943da875bb1b3b4f65146a41f2d0b265e0147c3fbd1a4710014b67657bd9 Selected ACT manifest: 4467cb11d5dcd466e1e10f8f47b4f98a8ecb6fab8a1608a1b315d76482490aec Kinematics: accb8aaab1e37e2f8bbb1f7de034e9206b81562e0e9879522c9e8a81d2e6e92a Supporting input SHA256 values: ancestor_freeze: 218755320ae874ef9ebc7e9aa6c9025135f0045e79f4015e661563648c466d89 concurrency_review: af9f641df6a65ba2f3de70892498cb89b53b86b8b343152466145f3d96917439 concurrency_review_image/review_images/LMDrillLift_overview_rgb.png: 5c45ab39dbc36a08d97a102b20aca9d79a88442427e0d4eafc2bb778f6e7e42b concurrency_review_image/review_images/LMDrillLift_rgb.png: a900fc7c23c8623b06cf436231b6f14d3d47897e30ddcb2570ee20a833a4ca70 concurrency_review_image/review_images/LMPushShelfForward_overview_rgb.png: 76470f6825b187671ddf318f0cf71dd8706027782b3e8e3884c0c96e718acc68 concurrency_review_image/review_images/LMPushShelfForward_rgb.png: ebeb5b90a4264ab00ba69959082a59316a55fb0d2b9d18c82e85dbcd049266c3 concurrency_supplement: 8e624357f086c514bf92c6b514e91c70ca51543913b5877cf7eb2b9ff23cce8a concurrency_validation: 8214b7a9fba1f5c12bac79b72d021a4f6722972a3495ebee2b7d139a00646e7f kinematics: accb8aaab1e37e2f8bbb1f7de034e9206b81562e0e9879522c9e8a81d2e6e92a previous_serial_batch: 0b18aecc40f208a2258f311811e72fe3a85a29fa69f8ba517cd034662170892b protocol: e08c943da875bb1b3b4f65146a41f2d0b265e0147c3fbd1a4710014b67657bd9 resumed_scheduler_plan: aa52191462cb682c696b8554fb227066d2b057f5b7470b19cb03a3fcf3a77975 retention_receipt: e21646d73447017beab042125e26f9662f3a4d9144a786e79b9ae76cf2bc7780 runtime_freeze: 1b63b1d5495fd24fa8d9b558c81e4be971f3d7154d6ec769155bee5fa79a4658 runtime_lineage: 9a644bb75428a0b72cd0e8e9d16770de8c7172dc8640aab9be3b24d9a1281cc5 scheduling_transition: f0552158d4079fc30a38e609138921f26d077aab3038c1913471e2e8336fc2c0 score_10000: 1e9c1dda2b824753953f4147f832c30ad40ed7da3a75da01fbc873a5982ceace score_20000: e04d697acf4bf2d61e73271cef9a524f5d64462b4cb36f5b6a35ed1bad4b3a84 score_30000: bd50a2c7dd59440add887c8676aa970aad8a1cf9ef092ee10835bae4a4648969 score_40000: d4a18afef6a0e74dc24d1648a95facd5b4b0b155b7c2fec51deb01f159f68abe selected_ensemble: 4467cb11d5dcd466e1e10f8f47b4f98a8ecb6fab8a1608a1b315d76482490aec serial_baseline: 3ac5889c3477a785c66fd017829a8e648cdfdab8d1c19fe7b359725fb56e8321 summary: 83da13f3cb002ba30a6bff50e485de9d4659140e26375450e2534fccb7d852d6 task_catalog: 5a0d8c3a8c16b06e1ac038b488fcee09b49e89355b7c3989cb085e42bd2e16b1 task_scene_contract: 5abdad8cdd01d7ed0bae78b14c124ce9d50d783263ee29fe01f91d99d8998d19 training_notes: 219bbdf6649fc0997c4bf04adb5e82f1f72c578d10fbfa98799d80feefd4befc training_statistics: 915772099a1949021a4e2ef577f166d9ab1723c95a246a54b35b56dc61aeeb58 v1_archive_receipt: 150fe06d086b158091dddcbe0e270440b2d1a23baf12c9dee2a6a90745ba92a7 v1_freeze: ed6450f48e3b2da2a9b2c7755c2d50dd331d98f0b175c31f5b5e1a29e7cb032f Primary sources: https://github.com/NVlabs/G1-Factory https://huggingface.co/datasets/nvidia/G1-Factory-Benchmark https://huggingface.co/nvidia/G1-Factory-GR00T-1.7 The companion report receipt binds this text to its exact inputs and generating source.