02Two research lanes · research only

Hands at work. People on foot.

Cr4dle plans two separate research lanes of data from consented people. Lane one is for hand work. Lane two is for people who walk among other consented people. No lane has a recorded session. Each lane has its own bets and decision records, and both lanes use the episode audit block of ADR-0054. The lane-two consent scheme is planned (ADR-0138 item 12.5). Cr4dle is a non-commercial research programme (ADR-0136).
01The two lanes

Two lanes.
No shared session.

A lane-two session records no hand-work task, and a lane-one session records no walk encounter. Each bet below shows its status and its kill date from horizons/bets.yaml.

  1. Hand work

    Lane one · research

    Consented people do physical work with their hands. Stand-in forearm bands record sEMG features, a calibrated label gives contact force, a head camera records egocentric video, and Neon glasses add gaze when they are worn.

    Cr4dle has recorded no lane-one session. Cr4dle holds no force data and no gaze data that it recorded from a person.

    Bets
    • contact-force-wrist-semg: funded, kill date
    • natural-gaze-target: funded, kill date
    • nail-force-module: proposed, kill date
    • contact-truth-audit: proposed, kill date
    • insole-contact-module: proposed, kill date
    • skill-capsule-learning: proposed, kill date
    • measured-vs-predicted-gaze: proposed, kill date
    • world-model-evaluator-force-ablation: proposed, kill date
    • semg-effort-from-video: proposed, kill date
    • reach-target-labels: proposed, kill date
    • force-conditioned-video-world-model: proposed, kill date
    • force-as-training-time-label: proposed, kill date
    • open-corpus-semg-force-stack: proposed, kill date
    • wm-human-conditioned-latent: proposed, kill date
    • human-model-paired-corpus: proposed, kill date
    • video-contact-lift: proposed, kill date
    • capsule-to-humanoid-retarget: proposed, kill date
    Decision records
    • ADR-0045: accepted (docs/decisions/0045-physical-intelligence-learned-from-humans.md)
    • ADR-0046: accepted (docs/decisions/0046-channel-taxonomy-with-label-validity.md)
    • ADR-0052: accepted (docs/decisions/0052-ember-capture-band-and-device-roles.md)
    • ADR-0053: accepted (docs/decisions/0053-acceptance-test-restated-after-verification-sweep.md)
    • ADR-0054: accepted (docs/decisions/0054-episode-audit-block.md)
    • ADR-0055: accepted (docs/decisions/0055-label-validity-v2-truth-in-labelling.md)
    • ADR-0120: accepted (docs/decisions/0120-simulated-demonstration-regime-for-the-policy-ablation.md)
    • ADR-0132: proposed (docs/decisions/0132-contact-truth-from-bare-hands.md)

  2. Human social navigation

    Lane two · research

    Consented people walk among other consented people in staged sessions. The record is human-side: gaze, scene video and head IMU, with body IMU planned. Lane two uses no sEMG.

    Cr4dle has recorded no lane-two encounter. The model trains only on synthetic encounters.

    Bets
    • pedestrian-gaze-on-robot: proposed, kill date
    • human-social-navigation-lane: proposed, kill date
    Decision records
    • ADR-0054: accepted (docs/decisions/0054-episode-audit-block.md)
    • ADR-0138: proposed (docs/decisions/0138-human-social-navigation-is-research-lane-two.md)
    • ADR-0139: proposed (docs/decisions/0139-the-social-navigation-encounter-protocol.md)
02Side by side

Each part.
Each state.

Planned marks a part that is not in the repository. A simulator result carries evidence_kind: simulation, and it is evidence about the code, not about people.

The two research lanes, part by part. Each status, kill date and evidence kind comes from the repository when the site builds. An equipment row with the status simulated has a driver and no bought unit.
PartLane one: Hand workLane two: Human social navigation
What is recorded
  • Forearm sEMG and IMU features from stand-in bands, still to be bought. Ember, the wrist band, is to replace them after its build gate. Its build gate is Q2 2027, and the build needs three paid collections on bought bands, so it waits on the decision that replaces ADR-0136.
  • Contact force: a label from the sEMG features, calibrated against a fingertip-force rig.
  • Egocentric video from a head camera, the default egocentric source. The video is redacted on the hub before it leaves.
  • Hand pose, derived from the head-camera video on the hub. Head pose: planned.
  • Gaze and head IMU from Neon glasses, a device to buy, when the glasses are worn. The kit adds them when a buyer names gaze (a rule for a later phase, ADR-0136).
  • A camera-side view from a fixed RGB-D camera. The video is redacted before it leaves, and depth stays on the device.
  • Ship reads in common/registry/taxonomy/capabilities.yaml: contact_force, gaze_target, hand_pose, head_pose, demonstrator_state.
  • Equipment row mindrove-armband-2: status simulated.
  • Equipment row oymotion-gforcepro-plus: status simulated.
  • Equipment row fingertip-force: status simulated.
  • Equipment row uvc-head-camera: status simulated.
  • Equipment row pupil-labs-neon: status simulated.
  • Equipment row exo-rgbd-camera: status simulated.
  • Equipment row led-flash-sync: status simulated.
  • Gaze, scene video and head IMU from Neon glasses, a device to buy. The video is redacted before it leaves.
  • The gaze_target validity in common/registry/taxonomy/capabilities.yaml covers 30 to 60 cm and 1.3 m only. ADR-0138 item 6.4 marks a gaze value at a larger encounter distance as not validated. That marker is planned, because no code sets it.
  • Head pose: planned. No lane-two kit gives head pose.
  • Body IMU features: planned. No equipment row exists.
  • Insoles: planned, after the bet insole-contact-module passes.
  • Positions of consented walkers from a fixed tracker: planned. The tracks stay on the device until a decision record sets their egress.
  • Bystanders: a redacted count only.
  • Audio: none.
  • The Neon kit of the Lens family has its own bet, natural-gaze-target: funded, kill date 2026-10-19.
  • Equipment row pupil-labs-neon: status simulated.
  • Equipment row moticon-opengo: status candidate.
  • Equipment row led-flash-sync: status simulated.
Tasks and rubrics
  • The task and rubric schemas are in platform/data-platform/schemas/tasks.
  • The rubric validator refuses a criterion that grades a research_only read.
  • Example bundle psp-sponge-cup: spec status draft, regime by_hand, ship reads contact_force, hand_pose, head_pose, consent scheme robot-training-v1, rubric psp-sponge-cup-rubric with the criterion groups data_quality, safety, outcome, kinematic, contact, instruction. Planned: the ship read head_pose, because the uvc-head-camera row gives only stub values (StaticHeadPoseStub). It is illustrative and holds no measured value.
  • A walker goes from a start marker into a goal zone. A consented person crosses the route.
  • The rubric grades physical facts of the path.
  • Human path ratings: planned (ADR-0139 D9). A rating rates the path, never the person.
  • The rubric never grades gaze_on_robot or social_context. Both stay research_only columns.
  • Example bundle socnav-walk-encounter: spec status draft, regime walking, ship reads gaze_target, head_pose, consent scheme lane-two-walk-planned, rubric socnav-walk-encounter-rubric with the criterion groups data_quality, safety, outcome, instruction, kinematic. Planned: the ship read head_pose, because no kit row of the bundle gives its stream in common/registry/config/equipment.yaml, and the consent scheme lane-two-walk-planned, because governance/consent/schemes/lane-two-walk-planned.yaml does not exist. It is illustrative and holds no measured value.
Environment
  • A work bench with objects to pick, squeeze and place. An LED flash fixture gives the clock sync.
  • Simulator task Cr4dle/SO101-DeformablePickSqueezePlace-v0 in sdk/sim/tasks/registry.yaml: status smoke, twin fidelity declared-unmeasured, evidence_kind: simulation. A simulated twin of the policy ablation. A result on it is evidence about the code, not about a person or a robot.
  • Gate 2 runs on the SO-101 ablation rig (rig_id so101-ablation-rig-v0, hardware/rigs/so101-ablation-rig/README.md). Its arm row so101-arm in common/registry/config/equipment.yaml has the status candidate, and Cr4dle owns no robot arm.
  • Staged sessions in a closed area. Each person in the area is a consented participant: a wearer, a co-walker or a confederate who walks to a script.
  • Sessions in a public area start only after an edit of ADR-0138.
  • Simulator task Cr4dle/WalkEncounter-v0 in sdk/sim/tasks/registry.yaml: status smoke, twin fidelity declared-unmeasured, evidence_kind: simulation. The scripted walkers are moving discs on fixed rules. A simulated walker is not a person.
  • The hardware suite socnav-walk-hw-v0: counterparts scripted_confederate, layout staged_intersection_for_cross, paces casual, neutral and rush, density 0.0 background_counterparts_per_m2 and corridor width 3.0 m. The suite plans a control walk with no counterpart for each walker, route and pace. It counts toward no floor. The sync fixture is led-flash-sync, and socnav-rec-1 sets the clock-offset limit. The tier-one producer is planned.
Benchmark
  • Protocol loso_regression in benchmarks/harness.yaml (Gate 1, ADR-0053).
  • Protocol policy_ablation in benchmarks/harness.yaml (Gate 2, ADR-0053; version 3 and the demonstration regime by_hand, ADR-0120).
  • Gate 1 rule (ADR-0053 items 13 and 14): the model learns from all participants but one and is tested on the held-out participant, for each participant in turn, with at least 5 participants. A per-user calibration of not more than 15 minutes is recorded from the held-out participant. This is the gated arm, cross_subject_loso_calib15. The calibration rows of the held-out participant fit only the calibration of that participant and are never scored (rule L7 of benchmarks/leakage_rules.md). Calibration rows never enter the training pool. Below 10 participants, the interval is descriptive. The model must do better than each baseline, and it must meet each target: a pooled per-finger mean absolute error of not more than 1.5 N, a per-finger error of not more than thumb 1.5 N, index 1.5 N, middle 2.0 N, ring 2.5 N and little 2.5 N, an R² of not less than 0.6, an error relative to strength of not more than 8 %MVC and a dropout50 error of not more than 1.25 times the clean pooled error. The arms cross_subject_loso, cross_subject_loso_finetune15 and semg_plus_pose are printed, not gated.
  • Gate 2 rule (ADR-0053 items 2 to 5): a policy with force and a policy without force are trained on the same by-hand episodes and do the same tasks on the same objects. Each policy does at least 40 trials. Gate 2 passes when the lower bound of the one-sided confidence interval of the gain in task success rate is more than zero and the observed gain is at least 15 points, a pre-registered minimum observed effect, not a certified margin. Any other result is KILL. The result is published either way.
  • Suite policy-ablation-so101-v0 (benchmarks/suites/policy-ablation-so101-v0.yaml): evidence_kind hardware, demonstration regime by_hand, with at least 40 trials in each arm. The gated pair is force_aware and motion_only. The exploratory arms estimated_force and glove_force are never gated.
  • Protocol socnav_encounter in benchmarks/harness.yaml (ADR-0139).
  • The unit is one encounter. Each split holds out whole persons (subject_id), and their counterparts too.
  • The scenario grid of socnav_encounter: pass (180°), overtake (0°), cross (45°, 90° and 135°), doorway and group_pass. Each suite sets the density and the corridor width.
  • The metrics of socnav_encounter: collision_count, min_separation_m, path_irregularity_rad, path_length_ratio, time_to_goal_s, success, stop_count, unforced_stop_count and pose_valid_fraction.
  • Suite socnav-walk-v0 (benchmarks/suites/socnav-walk-v0.yaml): evidence_kind simulation on Cr4dle/WalkEncounter-v0, printed only.
  • Suite socnav-walk-hw-v0 (benchmarks/suites/socnav-walk-hw-v0.yaml): evidence_kind hardware. A cell with fewer than 60 encounters or fewer than 20 persons is descriptive (ADR-0139 D4 item 5). The floors count gaze-valid encounters and the persons who walked them. The person floor counts each person with at least one gaze-valid encounter in the cell. The volume clause of the bet human-social-navigation-lane counts usable walkers, so it shares the number of the person floor of a decisive cell, not the unit.
  • Tier one scores recorded encounters offline. No encounter is recorded, so tier one has no result.
  • ADR-0139 sets no threshold, no PASS and no KILL before consented hardware encounters exist.
  • Each simulator number carries evidence_kind: simulation and is never release-eligible.
  • The simulator tier is a check of the pipeline under the frozen pre-registration socnav-pc-2 (research/projects/2026-10-socnav-walk-corpus/src/socnav_walk_corpus/preregistration.py). It has a null, an echo and a planted gaze control, an oracle arm and a calibration gate. Each reading has evidence_kind: simulation. A failed sensitivity part makes a tier-one stop reading inconclusive, and a failed specificity part makes each tier-one reading inconclusive. The check never stops lane two by itself (ADR-0138 item 9.7).
  • The stop rule of lane two is the kill criterion of the bet human-social-navigation-lane in horizons/bets.yaml.
Datasets
  • Own dataset own-2026-10-force-pilot-v1 in datasets/REGISTRY.yaml: status proposed.
  • External corpora hyser, putemg-force, ninapro-db2 and forceband-release in datasets/external/register.yaml: each has status candidate, confidence UNVERIFIED, intended_use eval-only and no licence_read.
  • A corpus stays evaluation-only until a person records licence_read (ADR-0086 item 3).
The episode format and a demo card on the datasets page
  • Lane two has no own dataset row in datasets/REGISTRY.yaml.
  • External corpora thor-magni, gazenav, musohu, egotraj, egowalk, egonav, vedb, egobody, aria-everyday-activities, urbanego, lookout-and, egocampus and nymeria in datasets/external/register.yaml: each has status candidate, confidence UNVERIFIED, intended_use eval-only and no licence_read.
  • Before counsel answers ADR-0138 item 11.6, no lane-two arm trains or runs on recorded people, an external walking corpus included.
  • After that answer, a corpus stays evaluation-only until a person records licence_read (ADR-0086 item 3).
Model
  • The force model reads sEMG and IMU features (SemgFeatureFrame) and writes a per-finger force with its calibration id (FingerForceEstimate). Its registered record has evidence_kind synthetic and the stage candidate, so it is not a released model.
  • No model is trained on a session that Cr4dle recorded, and Cr4dle keeps no trained weights.
  • A research project decides each bet.
  • Projects of the bet contact-force-wrist-semg: 2026-10-forceband-replication (active).
  • Projects of the bet natural-gaze-target: 2026-10-gaze-conditioned-il-replication (active).
  • Projects of the bet nail-force-module: 2026-10-nail-force-module (proposed).
  • Projects of the bet contact-truth-audit: 2026-10-contact-truth-audit (proposed).
  • Projects of the bet insole-contact-module: 2026-10-insole-contact-module (proposed).
  • Projects of the bet skill-capsule-learning: 2026-10-skill-capsule-learning (proposed).
  • Projects of the bet measured-vs-predicted-gaze: 2026-10-gaze-predicted-vs-measured (proposed).
  • Projects of the bet world-model-evaluator-force-ablation: 2026-11-imagined-ablation (proposed).
  • Projects of the bet semg-effort-from-video: 2026-12-semg-effort-from-video (proposed).
  • Projects of the bet reach-target-labels: 2026-12-reach-target-labels (proposed).
  • Projects of the bet force-conditioned-video-world-model: 2026-12-cosmos-force-conditioning-lora (proposed).
  • Projects of the bet force-as-training-time-label: 2026-10-sim-force-ablation-bc (proposed, evidence_kind simulation), 2026-12-force-as-training-label (proposed). The project 2026-10-sim-force-ablation-bc fits models on simulated episodes of the lane-one twin task, not on a session of a person.
  • Projects of the bet open-corpus-semg-force-stack: 2026-09-hyser-emg-force-loso (killed, evidence_kind external_corpus), 2026-10-electrophys-finger-decoder (proposed, evidence_kind external_corpus), 2026-10-emg-force-open-stack (proposed, evidence_kind external_corpus), 2026-10-force-model-stage1 (proposed), 2026-10-model-atlas-a (proposed).
  • Projects of the bet wm-human-conditioned-latent: 2027-01-latent-wm-channel-probe (proposed), 2027-02-persistent-state-evaluator (proposed).
  • Projects of the bet human-model-paired-corpus: 2027-06-camera-side-reach-target (proposed).
  • Projects of the bet video-contact-lift: 2027-06-video-contact-lift (proposed).
  • Projects of the bet capsule-to-humanoid-retarget: 2026-10-capsule-to-humanoid-retarget (proposed).
  • In the project 2026-10-force-model-stage1, S1 reads EMG after the scored row, so its verdict is for offline labelling only.
The results on the evidence page
  • A goal-conditioned waypoint policy: research/projects/2026-10-socnav-walk-corpus/src/socnav_walk_corpus/policy.py, simulator only, evidence_kind: simulation.
  • It trains only on synthetic encounters from the simulator.
  • Its inputs include head yaw and, in one arm, gaze yaw. Its output is the next waypoints of the walker agent only.
  • The decisive ablation is open loop. The closed-loop success of the policy is printed only (ADR-0138 items 9.1, 9.7 and 11.3).
  • The open-loop heads of research/projects/2026-10-socnav-walk-corpus/src/socnav_walk_corpus/openloop.py are head_only, head_plus_gaze and oracle. The pipeline of the path in force gates, and a recording record names that path before a session is recorded.
  • Projects of the bet pedestrian-gaze-on-robot: 2027-01-pedestrian-gaze (proposed).
  • Projects of the bet human-social-navigation-lane: 2026-10-socnav-walk-corpus (proposed).
  • In the project 2027-01-pedestrian-gaze, the bet records its encounters in lane-two walk sessions, but the project is not a lane-two project (ADR-0138 items 1.1 and 1.3).
03Limits

What each lane
never does.

These limits come from the decision records of each lane. A change to a limit needs an edit of its record.

Lane one: Hand work

  • Raw sEMG and raw gaze never leave the device. Video and poses leave only after redaction and under current consent.
  • Demonstrator state names only fatigue and physical states that are related to safety.
  • Neon output never trains the RGB gaze comparator.
  • A simulated scripted episode never gets to a dataset card, a model manifest or a release.
  • No product row names a research_only read.
  • No lane-one session records a walk encounter.

Lane two: Human social navigation

  • It uses no sEMG. A session uses no Ember band and no stand-in band.
  • No model forecasts the next action, path or goal of a bystander.
  • No arm forecasts the next action of a counterpart. The output of an arm is the path of the subject or a physical fact of the recording (ADR-0139 D6 items 4 and 5).
  • No metric uses a position, a pose or a track of a bystander.
  • No track, position, path, gaze target or identity of a bystander is recorded or leaves the device.
  • The gaze of a confederate who walks to a script is never a label of natural gaze.
  • No column, label or model output names a state of mind of a person.
  • Gaze is an input only, never a model target. No Neon output trains the model.
  • No person is recorded before the records of ADR-0138 items 12.4 and 12.5 exist and the ruling of item 5.5 is made.
  • A measured head-yaw source on the walking kit is also a gate. The pre-registration socnav-rec-1 names that source as the input of the head_only arm, and it adds the other gates that apply before lane two records a session.
  • No model trains or runs on recorded people, an external walking corpus included, before counsel rules on the question of ADR-0138 item 11.6.
  • A lane-two session records no hand-work task.
  • No product row names lane two.
04Sources

The records
hold the facts.

The site reads these files when it builds. If this page and a record do not agree, the record is correct.

The human sensoriumThe evidenceWhat you receiveThe demo

The site does not read the map of the two lanes. It is a document for a reader: docs/architecture/collection-lanes.md