Add common arm interfaces and exact simulation skill installation

README.md

@@ -6,6 +6,16 @@
 
 [Recorded demo](https://huggingface.co/spaces/Dvidia/dvidia-training) · [Synthetic datasets](https://huggingface.co/datasets/Dvidia/dvidia-training-examples) · [Pilot models](https://huggingface.co/Dvidia/dvidia-training-pilot)
 
+**[Arm interfaces and installation v0.1](docs/arm-interfaces-v01.md).**
+The **0.1.3.dev3 preview** adds `dvidia-arm`: three explicit native simulation
+adapters, exact PiPER policy compatibility checks, separate install/run commands,
+and a data-only lift-planning gate. A fixed tracking study retains **54/54
+numerically valid attempts**, including **21/1,260 offline compute-budget
+exceedances**. These are mechanics checks, not task success or cross-arm transfer.
+[Benchmark](benchmarks/arm-interfaces-v01/README.md) ·
+[Install the preview](https://github.com/Dvidia-Inference/dvidia-training/releases/tag/arm-interfaces-v0.1-preview) ·
+[Hugging Face evidence](https://huggingface.co/datasets/Dvidia/arm-interfaces-v01).
+
 **[Build a reusable task environment](docs/task-environments-v02.md).**
 The **0.1.3.dev2 native preview** adds a local metric editor and
 `dvidia-environment`: multiple rigid objects, an open bin, explicit OpenUSD and

benchmarks/arm-interfaces-v01/README.md

@@ -0,0 +1,42 @@
+# DVIDIA arm interface tracking evidence v0.1
+
+Three explicit native simulation interfaces: **PiPER**, **SO101**, and **UR5e + Robotiq 2F-85**.
+
+This is a local CPU mechanics and interface benchmark. It contains no trained policy, object manipulation task, physical robot connection or hardware-latency qualification.
+
+## What was run
+
+- 54 fixed attempts: three profiles × hold / bounded arm perturbation / gripper cycle × 10 / 20 / 40 ms updates × two repeats.
+- Every attempt simulates 0.4 seconds at a native 1 ms timestep, without hidden settling or automatic retries.
+- 54/54 attempts were numerically valid, with zero native warnings; this is not a task-success count.
+- 21,654 recorded state rows and 1,260 issued command events are retained.
+- **21 of 1,260 offline control blocks exceeded their nominal compute budget**: PiPER 13, SO101 1, UR5e/Robotiq 7. These are unpaced computation diagnostics, not measured physical deadlines.
+- The independent read-only audit recomputed 648 RMS/max values with zero trace/source digest or command-alignment mismatches.
+
+All cases and repeats are retained. Repeats use identical deterministic initial states and trajectories for timing variability; they do not establish generalization, confidence intervals or statistically independent task trials. No profile or update period was selected automatically.
+
+## Accuracy and speed mean specific things
+
+Arm tracking is measured in radians. PiPER gripper tracking uses its single-finger slide command in metres, SO101 uses a moving-jaw hinge target in radians, and Robotiq uses its native 0–255 servo coordinate. These are reported separately.
+
+Robotiq feedback is derived from the pinned affine servo model as `-(bias0 + bias1 * actuator_length) / gain`. It is not an aperture, force or tactile measurement. Raw state and the mapping parameters are retained.
+
+Tracking errors compare observed state with both the held issued target and the continuous reference at the observation time. Terminal-link position/orientation errors use native forward kinematics at the desired arm configuration; the link origin is not a calibrated fingertip tool centre point.
+
+`compile_and_initialize_wall_s` includes pinned asset verification, model compilation and initialization. `control_loop_compute_s` includes the complete control block. **`native_step_compute_s` measures the full observed-step wrapper: native stepping, forward update, state checks/snapshot and reference FK. It is not pure physics-engine execution time.** Trace serialization is outside the loop; total attempt timing stops before writing the final result receipt.
+
+The run used Python 3.12.9, MuJoCo 3.15.0 and NumPy 2.5.3 on Darwin arm64. The machine was an interactive shared host, not an isolated real-time test rig. Upstream gravity compensation and mechanics differ by profile. Do not interpret this table as a hardware quality ranking, minimum-system specification or sensor-to-actuator latency measurement.
+
+## Read the evidence
+
+[Interface and installation guide](../../docs/arm-interfaces-v01.md)
+
+- [Bounded summary](summary.json): all 27 condition summaries, settings, source hashes and timing scope.
+- [Independent audit](audit.json) and [original audit source](audit.py).
+- [Prepared artifact identities](publication-artifacts.json): archive and full inventory hashes.
+
+The companion evidence ZIP retains every raw trace; large traces are excluded from Git. The portable verifier is inside the ZIP. This directory records prepared artifacts and does not itself claim a completed publication.
+
+Archive: `arm-interfaces-v01-release-20261031.zip` · 11,159,124 bytes · SHA-256 `48200eb5c647badff7db5ee14272571b93d9c31003d9f5f92dd6e28d062b272b`.
+
+Original DVIDIA source and synthetic records: MIT. PiPER simulation assets: MIT; SO101: Apache-2.0; UR5e: BSD-3-Clause; Robotiq 2F-85: BSD-2-Clause. Full upstream notices are preserved in the companion archive.

benchmarks/arm-interfaces-v01/audit.json

@@ -0,0 +1,12 @@
+{
+  "schema": "dvidia.arm-interface-audit/v1",
+  "report_sha256": "5dd90299bd17515aee578631b46e3146c07ad18909eea838153d37bc4e9b1c5d",
+  "attempts": 54,
+  "state_rows": 21654,
+  "issued_command_events": 1260,
+  "independent_metric_comparisons": 648,
+  "source_and_trace_digest_mismatches": 0,
+  "command_time_alignment_mismatches": 0,
+  "offline_control_budget_exceedances": 21,
+  "scope": "Read-only numerical/digest audit; not a repeat run or physical validation."
+}

benchmarks/arm-interfaces-v01/audit.py

@@ -0,0 +1,71 @@
+"""Read-only independent audit of the fixed arm-interface tracking run."""
+from hashlib import sha256
+import json
+import math
+from pathlib import Path
+
+root=Path('runs/arm-interfaces-v01-20261031')
+report=json.loads((root/'report.json').read_text())
+protocol_bytes=(root/'protocol.json').read_bytes()
+assert sha256(protocol_bytes).hexdigest()==report['protocol']['sha256']
+protocol=json.loads(protocol_bytes)
+assert len(protocol['schedule'])==54==len(report['attempts'])
+assert all(sha256((root/'source'/f['name']).read_bytes()).hexdigest()==f['sha256'] for f in protocol['source_inventory'])
+assert len({a['id'] for a in report['attempts']})==54
+rows_total=commands_total=metric_comparisons=0
+for scheduled,result in zip(protocol['schedule'],report['attempts']):
+    assert all(result[k]==v for k,v in scheduled.items())
+    folder=root/result['id']
+    assert json.loads((folder/'result.json').read_text())==result
+    assert json.loads((folder/'started.json').read_text())['attempt']==scheduled
+    raw=(folder/'trace.json').read_bytes()
+    assert sha256(raw).hexdigest()==result['trace']['sha256'] and len(raw)==result['trace']['bytes']
+    trace=json.loads(raw)
+    rows,commands=trace['rows'],trace['commands']
+    assert result['valid'] and not result['error'] and not result['warnings'] and not result['warning_counters']
+    assert len(rows)==401 and len(commands)==round(.4/result['control_period_s'])
+    assert rows[-1]['elapsed_s']==.4 and result['simulated_seconds']==.4
+    substeps=round(result['control_period_s']/.001)
+    arm_indices=result['contract']['arm_action_indices']
+    gripper=result['gripper_coordinate']
+    series={key:[] for key in result['metrics'] if key!='state_rows'}
+    for index,row in enumerate(rows):
+        assert math.isclose(row['elapsed_s'],index*.001,abs_tol=1e-12)
+        assert math.isclose(row['native_time_s'],row['elapsed_s'],abs_tol=1e-10)
+        if index==0: continue
+        held=commands[(index-1)//substeps]['values_native']
+        assert held==row['applied_command_native']
+        reference=row['continuous_reference_command_native']
+        # Compiled interfaces have direct scalar arm joints at these qpos addresses.
+        actual=[row['qpos_native'][i] for i in arm_indices]
+        series['arm_held_target'].extend(value-held[i] for value,i in zip(actual,arm_indices))
+        series['arm_continuous_reference'].extend(value-reference[i] for value,i in zip(actual,arm_indices))
+        jaw=-(gripper['bias0']+gripper['bias1']*row['gripper_actuator_length_native'])/gripper['gain']
+        series['gripper_held_target'].append(jaw-held[gripper['action_index']])
+        series['gripper_continuous_reference'].append(jaw-reference[gripper['action_index']])
+        xyz,goal=row['terminal_link_position_m'],row['reference_terminal_link_position_m']
+        series['terminal_position_continuous_reference'].append(math.sqrt(sum((a-b)**2 for a,b in zip(xyz,goal))))
+        quat,goalquat=row['terminal_link_quaternion_wxyz'],row['reference_terminal_link_quaternion_wxyz']
+        dot=min(1.,max(0.,abs(sum(a*b for a,b in zip(quat,goalquat)))))
+        series['terminal_orientation_continuous_reference'].append(2*math.acos(dot))
+    for key,values in series.items():
+        measured=result['metrics'][key]
+        assert measured['count']==len(values)
+        rms=math.sqrt(sum(v*v for v in values)/len(values))
+        assert math.isclose(measured['rms'],rms,abs_tol=1e-10,rel_tol=1e-10),(result['id'],key,measured['rms'],rms)
+        assert math.isclose(measured['max_abs'],max(map(abs,values)),abs_tol=1e-10,rel_tol=1e-10)
+        metric_comparisons+=2
+    assert result['metrics']['arm_continuous_reference']['unit']=='rad'
+    assert result['metrics']['gripper_continuous_reference']['unit']==gripper['unit']
+    timing=result['timing']
+    assert len(timing['native_steps_s'])==400 and len(timing['control_intervals_s'])==len(commands)
+    assert timing['offline_control_budget_exceedances']==sum(v>result['control_period_s'] for v in timing['control_intervals_s'])
+    rows_total+=len(rows);commands_total+=len(commands)
+identity=sha256((root/'report.json').read_bytes()).hexdigest()
+output={'schema':'dvidia.arm-interface-audit/v1','report_sha256':identity,'attempts':54,
+        'state_rows':rows_total,'issued_command_events':commands_total,'independent_metric_comparisons':metric_comparisons,
+        'source_and_trace_digest_mismatches':0,'command_time_alignment_mismatches':0,
+        'offline_control_budget_exceedances':sum(a['timing']['offline_control_budget_exceedances'] for a in report['attempts']),
+        'scope':'Read-only numerical/digest audit; not a repeat run or physical validation.'}
+Path('runs/arm-interfaces-v01-audit-20261031.json').write_text(json.dumps(output,indent=2)+'\n')
+print(json.dumps(output,indent=2))