{ "@context": "https://schema.certanix.eu/v1/evidence-context.jsonld", "@type": "EvidencePack", "schemaVersion": "1.14", "id": "evp:embodied:wp2-aloha-transfer-cube-v1-act-pergroup-s0-n40-k8-release-b-v3:2026-11-08T17:46:25Z", "subject": { "tenant": "tenant-certanix-demo", "device_manifest": "none", "ai_policy_uri": "hf://lerobot/act_aloha_sim_transfer_cube_human@bf09d3a9c6fd382bbcae1167297f4ec88a38f650", "policy_hash": "sha256:d9f89098a0ec86c4259cf6b664872fb8b22f3734863cf9024ad4884c9347d191" }, "campaign": { "campaign_id": "wp2-aloha-transfer-cube-v1-act-pergroup-s0-n40-k8-release-b-v3", "scenarios_executed": 8, "environments_per_scenario": 1, "total_simulation_hours": 0.014166666666666666, "assurance_target": null, "scenario_pass_rate": 0.625, "adversarial_robustness_score": null, "tail_probability_estimate": null, "explainability_coverage": null, "dataset_diversity_score": null, "dataset_diversity_uplift": null }, "scenario_results": [ { "scenario_id": "aloha-transfer-cube-v1-s0-c021", "name": "AlohaTransferCube-v0", "status": "failed", "environments_executed": 1, "metrics": { "contact_count_max": 1.0, "sim_time_s": 8.0, "steps_executed": 400.0, "task_reward_max": 0.0 }, "anomalies": [], "seed": 21, "deterministic_hash": "ef1c3514cf0155e34af2dbff16a331a1d67c70f3f8b61eded08c9f046570a2c8" }, { "scenario_id": "aloha-transfer-cube-v1-s0-c027", "name": "AlohaTransferCube-v0", "status": "failed", "environments_executed": 1, "metrics": { "contact_count_max": 1.0, "sim_time_s": 8.0, "steps_executed": 400.0, "task_reward_max": 0.0 }, "anomalies": [], "seed": 27, "deterministic_hash": "accecc8ebda992c28e86e88e1307c9bd65a7379950093848adcf8a50d556d0b5" }, { "scenario_id": "aloha-transfer-cube-v1-s0-c000", "name": "AlohaTransferCube-v0", "status": "passed", "environments_executed": 1, "metrics": { "contact_count_max": 4.0, "sim_time_s": 5.72, "steps_executed": 286.0, "task_reward_max": 4.0 }, "anomalies": [], "seed": 0, "deterministic_hash": "b5b8ad78fa85e4e74b5b8320c49a77bb5aa355eaf80b1fa7d7137900b0b3dd70" }, { "scenario_id": "aloha-transfer-cube-v1-s0-c011", "name": "AlohaTransferCube-v0", "status": "passed", "environments_executed": 1, "metrics": { "contact_count_max": 4.0, "sim_time_s": 5.68, "steps_executed": 284.0, "task_reward_max": 4.0 }, "anomalies": [], "seed": 11, "deterministic_hash": "a1019a49c79f8232744b8b4defa1d88e7dc771d67ac424c35f3d26deb9022004" }, { "scenario_id": "aloha-transfer-cube-v1-s0-c020", "name": "AlohaTransferCube-v0", "status": "failed", "environments_executed": 1, "metrics": { "contact_count_max": 3.0, "sim_time_s": 8.0, "steps_executed": 400.0, "task_reward_max": 2.0 }, "anomalies": [], "seed": 20, "deterministic_hash": "336b922cba38d9476b11f96a14cfc98cfd2ab186c5a86a80bc9d642dc83d2be5" }, { "scenario_id": "aloha-transfer-cube-v1-s0-c031", "name": "AlohaTransferCube-v0", "status": "passed", "environments_executed": 1, "metrics": { "contact_count_max": 3.0, "sim_time_s": 4.32, "steps_executed": 216.0, "task_reward_max": 4.0 }, "anomalies": [], "seed": 31, "deterministic_hash": "639eb651bcbdb1a4e4d568f92f5471702264aec9d9dbbe4f3a24c40e7f0bab75" }, { "scenario_id": "aloha-transfer-cube-v1-s0-c023", "name": "AlohaTransferCube-v0", "status": "passed", "environments_executed": 1, "metrics": { "contact_count_max": 4.0, "sim_time_s": 5.54, "steps_executed": 277.0, "task_reward_max": 4.0 }, "anomalies": [], "seed": 23, "deterministic_hash": "b7c7f7cbe41eba8acc35f1591a844e86307b29e86fe2a3d6a826373081a5c770" }, { "scenario_id": "aloha-transfer-cube-v1-s0-c034", "name": "AlohaTransferCube-v0", "status": "passed", "environments_executed": 1, "metrics": { "contact_count_max": 4.0, "sim_time_s": 5.74, "steps_executed": 287.0, "task_reward_max": 4.0 }, "anomalies": [], "seed": 34, "deterministic_hash": "f1f72f77f165ffd0eda39148b73397938bbcc904e4bdef2207d4d0b393481be3" } ], "shap_records": [], "compliance_mappings": { "ai_act_annex_iv": {}, "easa_l2_ml": {}, "do_178c": null, "annex_iv_clauses": [], "cra_annex_i_clauses": [], "regulation_clauses": [ { "framework": "machinery-regulation", "instrument": "Regulation (EU) 2023/1230 (Machinery Regulation), OJ L 165, 29.6.2023, as corrected by OJ L 169, 4.7.2023", "clause": "Annex IV Part A (g)", "clause_text_ref": { "file": "references/regulation-clauses.md#machinery-annex-iv-a-g", "legal": "Regulation (EU) 2023/1230 (CELEX 32023R1230), Annex IV, Part A, point (g)" }, "relation": "relates-to", "route": "Internal production control (module A, Art. 25(4) and Annex VI) against the Annex III essential health and safety requirements; no notified body. Annex I, Part A, points 5 and 6 concern machine learning that both ensures a safety function and is fully or partially self-evolving (recital 55); this is a frozen task policy. Applies from 20 January 2027 (Art. 54, as corrected); Directive 2006/42/EC until then.", "why": "Results of tests on the control logic: the task, the success definition signed before the results, the seeds, per-scenario verdicts and the trace.", "not_shown": "Which essential health and safety requirement each test verifies (the manufacturer's risk assessment, Part A (b), makes that link); tests on the physical machine.", "evidence_artifacts": [ { "type": "embodied-trace", "content_id": "sha256:cd89fe69dda34bbae769fa00f41974a7094e4344e272d660effd85aa5c6e1b6b" }, { "type": "parameter-coverage-map", "content_id": "sha256:c8a4470e825572deb8da64f52cd844e76e3265f96fb3f26fd411929e6fc11c04" } ], "source": null }, { "framework": "machinery-regulation", "instrument": "Regulation (EU) 2023/1230 (Machinery Regulation), OJ L 165, 29.6.2023, as corrected by OJ L 169, 4.7.2023", "clause": "Annex III 1.1.2(a)", "clause_text_ref": { "file": "references/regulation-clauses.md#machinery-annex-iii-1-1-2-a", "legal": "Regulation (EU) 2023/1230 (CELEX 32023R1230), Annex III, section 1.1.2(a)" }, "relation": "relates-to", "route": "Internal production control (module A, Art. 25(4) and Annex VI) against the Annex III essential health and safety requirements; no notified body. Annex I, Part A, points 5 and 6 concern machine learning that both ensures a safety function and is fully or partially self-evolving (recital 55); this is a frozen task policy. Applies from 20 January 2027 (Art. 54, as corrected); Directive 2006/42/EC until then.", "why": "Fitness for its function: measured task success under the selected variations.", "not_shown": "That the machine can be operated without putting persons at risk; reasonably foreseeable misuse.", "evidence_artifacts": [ { "type": "embodied-trace", "content_id": "sha256:cd89fe69dda34bbae769fa00f41974a7094e4344e272d660effd85aa5c6e1b6b" }, { "type": "parameter-coverage-map", "content_id": "sha256:c8a4470e825572deb8da64f52cd844e76e3265f96fb3f26fd411929e6fc11c04" } ], "source": null }, { "framework": "machinery-regulation", "instrument": "Regulation (EU) 2023/1230 (Machinery Regulation), OJ L 165, 29.6.2023, as corrected by OJ L 169, 4.7.2023", "clause": "Annex III 1.2.1, first paragraph and (c)", "clause_text_ref": { "file": "references/regulation-clauses.md#machinery-annex-iii-1-2-1", "legal": "Regulation (EU) 2023/1230 (CELEX 32023R1230), Annex III, section 1.2.1, first paragraph and point (c)" }, "relation": "relates-to", "route": "Internal production control (module A, Art. 25(4) and Annex VI) against the Annex III essential health and safety requirements; no notified body. Annex I, Part A, points 5 and 6 concern machine learning that both ensures a safety function and is fully or partially self-evolving (recital 55); this is a frozen task policy. Applies from 20 January 2027 (Art. 54, as corrected); Directive 2006/42/EC until then.", "why": "The behaviour of the control logic across the tested variations, including the failing cases.", "not_shown": "That an error in the control logic does not lead to a hazardous situation: the campaign scores task success, not hazards.", "evidence_artifacts": [ { "type": "embodied-trace", "content_id": "sha256:cd89fe69dda34bbae769fa00f41974a7094e4344e272d660effd85aa5c6e1b6b" }, { "type": "parameter-coverage-map", "content_id": "sha256:c8a4470e825572deb8da64f52cd844e76e3265f96fb3f26fd411929e6fc11c04" } ], "source": null }, { "framework": "machinery-regulation", "instrument": "Regulation (EU) 2023/1230 (Machinery Regulation), OJ L 165, 29.6.2023, as corrected by OJ L 169, 4.7.2023", "clause": "Annex III 1.2.1(d)", "clause_text_ref": { "file": "references/regulation-clauses.md#machinery-annex-iii-1-2-1-d", "legal": "Regulation (EU) 2023/1230 (CELEX 32023R1230), Annex III, section 1.2.1, point (d)" }, "relation": "relates-to", "route": "Internal production control (module A, Art. 25(4) and Annex VI) against the Annex III essential health and safety requirements; no notified body. Annex I, Part A, points 5 and 6 concern machine learning that both ensures a safety function and is fully or partially self-evolving (recital 55); this is a frozen task policy. Applies from 20 January 2027 (Art. 54, as corrected); Directive 2006/42/EC until then.", "why": "The exact policy tested is pinned (weights sha256, and the weights actually loaded); a different policy is a different pack.", "not_shown": "That the machine prevents modification of its settings or rules in use; the limits of the safety functions.", "evidence_artifacts": [], "source": null }, { "framework": "ai-act", "instrument": "Regulation (EU) 2024/1689 (Artificial Intelligence Act), consolidated text CELEX 02024R1689-20260727", "clause": "Annex IV 2(g)", "clause_text_ref": { "file": "references/regulation-clauses.md#ai-act-annex-iv-2g", "legal": "Regulation (EU) 2024/1689 (CELEX 02024R1689-20260727), Annex IV, point 2(g)" }, "relation": "relates-to", "route": "Likely not high-risk: Art. 6(1)(b) needs a third-party conformity assessment under the Annex I legislation, and the Machinery route for this frozen task policy is internal production control (see the Machinery entries). The Machinery Regulation is in Annex I, Section B (item 21, as amended by Regulation (EU) 2026/1744), where Art. 9, 10, 15 and Annex IV do not apply directly (Art. 2(2)). Related here as reference requirements only.", "why": "The testing procedure in simulation: the task, the success definition signed before the results, the seeds, per-scenario verdicts and the trace.", "not_shown": "Testing in real-world conditions (Art. 60, 60a); the gap between simulation and the real cell; test reports signed by the provider's responsible persons.", "evidence_artifacts": [ { "type": "embodied-trace", "content_id": "sha256:cd89fe69dda34bbae769fa00f41974a7094e4344e272d660effd85aa5c6e1b6b" }, { "type": "parameter-coverage-map", "content_id": "sha256:c8a4470e825572deb8da64f52cd844e76e3265f96fb3f26fd411929e6fc11c04" } ], "source": null }, { "framework": "ai-act", "instrument": "Regulation (EU) 2024/1689 (Artificial Intelligence Act), consolidated text CELEX 02024R1689-20260727", "clause": "Annex IV 1(c)", "clause_text_ref": { "file": "references/regulation-clauses.md#ai-act-annex-iv-1c", "legal": "Regulation (EU) 2024/1689 (CELEX 02024R1689-20260727), Annex IV, point 1(c)" }, "relation": "relates-to", "route": "Likely not high-risk: Art. 6(1)(b) needs a third-party conformity assessment under the Annex I legislation, and the Machinery route for this frozen task policy is internal production control (see the Machinery entries). The Machinery Regulation is in Annex I, Section B (item 21, as amended by Regulation (EU) 2026/1744), where Art. 9, 10, 15 and Annex IV do not apply directly (Art. 2(2)). Related here as reference requirements only.", "why": "Software versions: policy repository and revision, weights sha256, harness pins and component versions.", "not_shown": "Firmware on the physical robot; requirements related to version updates.", "evidence_artifacts": [], "source": null }, { "framework": "ai-act", "instrument": "Regulation (EU) 2024/1689 (Artificial Intelligence Act), consolidated text CELEX 02024R1689-20260727", "clause": "Annex IV 3", "clause_text_ref": { "file": "references/regulation-clauses.md#ai-act-annex-iv-3", "legal": "Regulation (EU) 2024/1689 (CELEX 02024R1689-20260727), Annex IV, point 3" }, "relation": "relates-to", "route": "Likely not high-risk: Art. 6(1)(b) needs a third-party conformity assessment under the Annex I legislation, and the Machinery route for this frozen task policy is internal production control (see the Machinery entries). The Machinery Regulation is in Annex I, Section B (item 21, as amended by Regulation (EU) 2026/1744), where Art. 9, 10, 15 and Annex IV do not apply directly (Art. 2(2)). Related here as reference requirements only.", "why": "Measured task success and the failing scenarios, under the tested variations.", "not_shown": "The safety of persons near the arm; foreseeable unintended outcomes in the real cell.", "evidence_artifacts": [ { "type": "embodied-trace", "content_id": "sha256:cd89fe69dda34bbae769fa00f41974a7094e4344e272d660effd85aa5c6e1b6b" }, { "type": "parameter-coverage-map", "content_id": "sha256:c8a4470e825572deb8da64f52cd844e76e3265f96fb3f26fd411929e6fc11c04" } ], "source": null }, { "framework": "ai-act", "instrument": "Regulation (EU) 2024/1689 (Artificial Intelligence Act), consolidated text CELEX 02024R1689-20260727", "clause": "Art. 9(8)", "clause_text_ref": { "file": "references/regulation-clauses.md#ai-act-art-9-8", "legal": "Regulation (EU) 2024/1689 (CELEX 02024R1689-20260727), Article 9(8)" }, "relation": "relates-to", "route": "Likely not high-risk: Art. 6(1)(b) needs a third-party conformity assessment under the Annex I legislation, and the Machinery route for this frozen task policy is internal production control (see the Machinery entries). The Machinery Regulation is in Annex I, Section B (item 21, as amended by Regulation (EU) 2026/1744), where Art. 9, 10, 15 and Annex IV do not apply directly (Art. 2(2)). Related here as reference requirements only.", "why": "Tested against a success criterion defined before the run.", "not_shown": "The risk management system; the Machinery risk assessment.", "evidence_artifacts": [ { "type": "embodied-trace", "content_id": "sha256:cd89fe69dda34bbae769fa00f41974a7094e4344e272d660effd85aa5c6e1b6b" } ], "source": null }, { "framework": "ai-act", "instrument": "Regulation (EU) 2024/1689 (Artificial Intelligence Act), consolidated text CELEX 02024R1689-20260727", "clause": "Art. 15(4)", "clause_text_ref": { "file": "references/regulation-clauses.md#ai-act-art-15-4", "legal": "Regulation (EU) 2024/1689 (CELEX 02024R1689-20260727), Article 15(4)" }, "relation": "relates-to", "route": "Likely not high-risk: Art. 6(1)(b) needs a third-party conformity assessment under the Annex I legislation, and the Machinery route for this frozen task policy is internal production control (see the Machinery entries). The Machinery Regulation is in Annex I, Section B (item 21, as amended by Regulation (EU) 2026/1744), where Art. 9, 10, 15 and Annex IV do not apply directly (Art. 2(2)). Related here as reference requirements only.", "why": "Resilience across the task-parameter variations the campaign selects (recorded in the coverage map).", "not_shown": "Fail-safe and redundancy; behaviour outside the simulated variations.", "evidence_artifacts": [ { "type": "embodied-trace", "content_id": "sha256:cd89fe69dda34bbae769fa00f41974a7094e4344e272d660effd85aa5c6e1b6b" }, { "type": "parameter-coverage-map", "content_id": "sha256:c8a4470e825572deb8da64f52cd844e76e3265f96fb3f26fd411929e6fc11c04" } ], "source": null } ], "frameworks_considered_not_related": [ { "framework": "cra", "reason": "The campaign tests task behaviour in simulation, not a cybersecurity property, and no SBOM or vulnerability scan is ingested." } ] }, "artefacts": [ { "type": "embodied-trace", "cid": "sha256:cd89fe69dda34bbae769fa00f41974a7094e4344e272d660effd85aa5c6e1b6b" }, { "type": "parameter-coverage-map", "cid": "sha256:c8a4470e825572deb8da64f52cd844e76e3265f96fb3f26fd411929e6fc11c04" } ], "ingested_artifacts": [], "provenance": { "platform_version": "certanix-beta@e9b999d5", "ran_at": "2026-11-08T17:46:25Z", "ran_on": null, "component_versions": { "mujoco": "3.3.6", "dm_control": "1.0.34", "gym-aloha": "0.1.4", "gymnasium": "1.3.0", "lerobot": "0.6.1", "torch": "2.11.0", "torchvision": "0.26.0", "numpy": "2.2.6", "PyOpenGL": "3.1.10", "glfw": "2.10.2" }, "quantum_job_ids": [], "quantum_job_devices": [], "quantum_beat": null, "vehicle": null, "selection_solver_class": null, "platform_family": "macos-arm64", "speed_fidelity": null, "twin_fidelity": null, "refusal_reason": null, "harness": { "harness_class": "sim-rigid-body-mujoco", "provider": "mujoco", "capabilities": [ "actuator_command", "contact_pairs", "joint_state", "joint_velocity", "object_pose", "physical_interaction", "rgb_camera" ], "pins_sha256": "a3a59a08b2fc8bf741d0c380ca8a2b10b7e259439801d2309b31f6fe72e814b8", "pins_path": "tools/mujoco-pin.v3.json" }, "task": { "task_id": "AlohaTransferCube-v0", "env_package": "gym-aloha", "env_revision": "gym-aloha 0.1.4, PyPI gym_aloha-0.1.4-py3-none-any.whl sha256 d9044290fbccddf0be4246b5287cf0eb6b9ddee545a3d222ce8d78c93ce7125e", "success_definition": "gym_aloha/env.py AlohaEnv.step: `terminated = is_success = reward == 4`. gym_aloha/tasks/sim.py TransferCubeTask.get_reward, from contact pairs: 1 = cube touches the right gripper finger; 2 = and not the table (lifted); 3 = cube touches the left gripper finger (attempted transfer); 4 = left finger and not the table (successful transfer). max_reward = 4." }, "policy": { "policy_class": "policy-learned-deterministic", "repo": "lerobot/act_aloha_sim_transfer_cube_human", "revision": "bf09d3a9c6fd382bbcae1167297f4ec88a38f650", "weights_sha256": "d9f89098a0ec86c4259cf6b664872fb8b22f3734863cf9024ad4884c9347d191", "loaded_weights_sha256": "63dd2851167c091e2dc2f2859f77f203291f06ddf93284c83fd932d8839544a8", "derivation": "config.json converted from LeRobot's 2024 format by tools/embodied_assets.py convert_2024_config: type set to \"act\"; input_shapes and output_shapes become input_features and output_features (observation.images.* VISUAL, observation.state STATE, action ACTION); the per-feature normalisation modes become normalization_mapping by feature type; temporal_ensemble_momentum null becomes temporal_ensemble_coeff null; every other field is copied unchanged; the weights are not touched (applied to revision bf09d3a9c6fd382bbcae1167297f4ec88a38f650, published 2024-07-16; converted config.json sha256 d23c453b02ffdc3b0403f5b1223c7b93de620278409c0f27d0d61d2c9c629c5c); then lerobot 0.6.1 `python -m lerobot.processor.migrate_policy_normalization --pretrained-path --output-dir `, run offline (HF_HUB_OFFLINE=1) on the published revision, which moves the normalisation statistics out of the weights into processor files; then the device fields in config.json and policy_preprocessor.json are set to \"cpu\", the device inference runs on. Without that second step the migration writes whatever accelerator the deriving host auto-selects in place of the published \"cuda\" (measured: \"mps\" on macos-arm64, \"cpu\" on linux-x86_64), so the derived files differed by platform", "inference_device": "cpu", "deterministic_algorithms": true, "num_threads": 1 }, "renderer": { "backend": "glfw", "gl_vendor": "Apple", "gl_renderer": "Apple M1 Max", "gl_version": "2.1 Metal - 90.5", "camera": "top", "width": 640, "height": 480 } }, "signature": { "algorithm": "ECDSA-P256", "key_source": "yubihsm-hardware", "signer_dn": null, "signature_b64": "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" } }