YAM scripted API control
passed
7 completed commands · 99.9 mm maximum lift · 1.0 mm placement error
Results across three arms
SO100, YAM, and SO101 comparison
Each row states the controller and evaluation population so the outcomes can be read in context. SO100 uses its own observations. YAM and SO101 use live policy outputs from the SO100 shadow scene. Their held out results follow controller corrections on seed 42001.
| Arm | Controller | Evaluation group | Trials | Successes | Failures | Evidence context |
|---|---|---|---|---|---|---|
| SO100 | V3 policy with the SO100 scene cameras and measured joints | Primary evaluation | 12 | 10 | 2 | 12 of 12 paired outcomes agree with direct control |
| YAM | Live V3 outputs from an SO100 shadow scene, mapped through FK and IK to target API commands | Fixed controller on held out seeds | 2 | 2 | 0 | Full attempt history: 3 successes, 1 failures; includes controller development and held out trials |
| SO101 | Live V3 outputs from an SO100 shadow scene, mapped through FK and IK to target API commands | Fixed controller on held out seeds | 2 | 2 | 0 | Full attempt history: 3 successes, 5 failures; includes controller development and held out trials |
Trained policy / SO100
Task outcomes
10/12API policy successes
10/12Direct MuJoCo reference successes
12/12Paired success agreement
55.2% to 95.3%Wilson 95% interval for API success
Success is the benchmark's ordered physical predicate: bilateral cube contact, lift, containment inside the green zone, open jaw release, and at least 0.5 simulated seconds settled. A completed HTTP request alone does not count as success. Seeds and full failure traces are retained. The confidence interval describes sampling uncertainty for these seeds only.
Paired outcome ledger
| Seed | API success | Direct success | API ticks | Direct ticks | API wall s | Direct wall s | Max action Δ | Max cube Δ m | API error |
|---|---|---|---|---|---|---|---|---|---|
| 42000 | yes | yes | 411 | 411 | 42.7 | 9.9 | 0.00000000 | 0.00000000 | Not recorded |
| 42001 | yes | yes | 401 | 401 | 41.8 | 9.4 | 0.00410157 | 0.00003489 | Not recorded |
| 42002 | yes | yes | 377 | 377 | 38.6 | 8.9 | 0.00000000 | 0.00000000 | Not recorded |
| 42003 | yes | yes | 338 | 337 | 34.9 | 7.1 | Not recorded | Not recorded | Not recorded |
| 42004 | no | no | 600 | 600 | 61.7 | 12.3 | 0.00000000 | 0.00000000 | Not recorded |
| 42005 | yes | yes | 396 | 396 | 40.6 | 8.1 | 0.00000000 | 0.00000000 | Not recorded |
| 42006 | yes | yes | 350 | 350 | 36.2 | 7.1 | 0.00000000 | 0.00000000 | Not recorded |
| 42007 | yes | yes | 347 | 346 | 35.6 | 7.2 | Not recorded | Not recorded | Not recorded |
| 42008 | no | no | 600 | 600 | 61.0 | 12.3 | 0.00000000 | 0.00000000 | Not recorded |
| 42009 | yes | yes | 348 | 348 | 35.0 | 7.2 | 0.00000000 | 0.00000000 | Not recorded |
| 42010 | yes | yes | 396 | 401 | 38.1 | 8.3 | Not recorded | Not recorded | Not recorded |
| 42011 | yes | yes | 353 | 353 | 35.8 | 7.3 | 0.01503906 | 0.00007006 | Not recorded |
Trajectory deltas compare corresponding recorded ticks only when both traces have the same length. Zero means the API reproduced the direct simulation numerically for that seed; a dash means comparison was unavailable.
Execution path
Latency and control
| Operation | Calls | Median ms | P95 ms | Worst ms |
|---|---|---|---|---|
| Atomic two camera observation | 4917 | 49.12 | 51.27 | 182.59 |
| Trained policy RPC | 621 | 227.24 | 239.29 | 258.25 |
| API control step + MuJoCo physics | 4917 | 15.14 | 16.45 | 115.17 |
Statistics pool every recorded call across API episodes; P95 uses the nearest rank definition. Latency is measured at the loopback HTTP client, including response parsing. Policy RPC time includes server computation and transport to the local policy server. API step latency includes validation, journaling, and MuJoCo integration. Physics is paused during policy inference in this experiment, so these timings do not establish a hard real time robot controller.
API execution audit
What the API did, call by call
The primary SO100 evaluation executed 4917 authenticated POST /v1/sessions/{id}/policy-step calls and 621 actual V3 policy inference RPCs across 12 sessions. Every control tick paired a fresh atomic camera plus joint observation with one absolute six joint target and returned measured MuJoCo state. Tokens and lease secrets are not printed here.
| Order | HTTP call | Payload / response |
|---|---|---|
| 1 | POST /v1/sessions | robot_id=so100, scene=groot_v3, manual mode, recording on; response has session ID and lease epoch. |
| 2, every tick | GET /v1/sessions/{id}/observation?include_rgb=true | Atomically returns front/wrist RGB, joint positions, sequence, object/task state. |
| 3, at chunk boundary | V3 policy RPC | Two live camera frames, measured joint state and task instruction; returns 40 predicted absolute six joint actions. |
| 4, every tick | POST /v1/sessions/{id}/policy-step | Lease header, fresh observation sequence, increasing command sequence and the current model produced joint target; returns measured state, contact/task metrics and simulation time. |
| 5 | DELETE /v1/sessions/{id}; GET /export | Close the session, then download the full per tick API journal. |
Actual first control tick, seed 42000
{
"observation_sequence": 1,
"command_sequence": 0,
"positions_rad": [
0.5325309634208679,
-1.3794503211975098,
1.328666090965271,
1.6192513704299927,
-2.1013333797454834,
0.699999988079071
],
"measured_positions_rad": [
0.5310142602164671,
-1.367295844510259,
1.3273086097611002,
1.625157377439205,
-2.101003908060749,
0.6999829452756653
],
"object_position_m_after_step": [
0.09933514190447397,
-0.22944205256490283,
0.0148922445562271
],
"task_success_after_step": false
}Full model/action/physics trace · Raw API episode journal · HTTP access log · Run contract
Exploratory follow up
Preserved failures and controller repair
The primary 12 seed result above is unchanged. We reran only its two failed seeds with different policy action commitment horizons; this selected seed follow up is not an unbiased new success rate. Commitment 16 rescued seed 42004 through both API and direct control (347 ticks); seed 42008 still failed. All failed trials and videos remain linked.
Per seed evidence
Policy driven pick and place
Seed 42000
SUCCESS411 policy ticks · 13.70 simulated seconds · 52 policy calls · 272 bilateral contact ticks · 72.2 mm peak cube center height · direct reference: success
Through authenticated API
Direct MuJoCo reference
Measured object path
Cube center height (mm)
API control step latency (ms)
Seed 42001
SUCCESS401 policy ticks · 13.37 simulated seconds · 51 policy calls · 272 bilateral contact ticks · 75.2 mm peak cube center height · direct reference: success
Through authenticated API
Direct MuJoCo reference
Measured object path
Cube center height (mm)
API control step latency (ms)
Seed 42002
SUCCESS377 policy ticks · 12.57 simulated seconds · 48 policy calls · 245 bilateral contact ticks · 74.6 mm peak cube center height · direct reference: success
Through authenticated API
Direct MuJoCo reference
Measured object path
Cube center height (mm)
API control step latency (ms)
Seed 42003
SUCCESS338 policy ticks · 11.27 simulated seconds · 43 policy calls · 202 bilateral contact ticks · 73.9 mm peak cube center height · direct reference: success
No video captured.
Measured object path
Cube center height (mm)
API control step latency (ms)
Seed 42004
FAILURE600 policy ticks · 20.00 simulated seconds · 75 policy calls · 489 bilateral contact ticks · 74.4 mm peak cube center height · direct reference: failure
No video captured.
Measured object path
Cube center height (mm)
API control step latency (ms)
Seed 42005
SUCCESS396 policy ticks · 13.20 simulated seconds · 50 policy calls · 260 bilateral contact ticks · 72.8 mm peak cube center height · direct reference: success
No video captured.
Measured object path
Cube center height (mm)
API control step latency (ms)
Seed 42006
SUCCESS350 policy ticks · 11.67 simulated seconds · 44 policy calls · 222 bilateral contact ticks · 73.8 mm peak cube center height · direct reference: success
No video captured.
Measured object path
Cube center height (mm)
API control step latency (ms)
Seed 42007
SUCCESS347 policy ticks · 11.57 simulated seconds · 44 policy calls · 222 bilateral contact ticks · 75.7 mm peak cube center height · direct reference: success
No video captured.
Measured object path
Cube center height (mm)
API control step latency (ms)
Seed 42008
FAILURE600 policy ticks · 20.00 simulated seconds · 75 policy calls · 447 bilateral contact ticks · 20.6 mm peak cube center height · direct reference: failure
No video captured.
Measured object path
Cube center height (mm)
API control step latency (ms)
Seed 42009
SUCCESS348 policy ticks · 11.60 simulated seconds · 44 policy calls · 221 bilateral contact ticks · 72.7 mm peak cube center height · direct reference: success
No video captured.
Measured object path
Cube center height (mm)
API control step latency (ms)
Seed 42010
SUCCESS396 policy ticks · 13.20 simulated seconds · 50 policy calls · 263 bilateral contact ticks · 73.5 mm peak cube center height · direct reference: success
No video captured.
Measured object path
Cube center height (mm)
API control step latency (ms)
Seed 42011
SUCCESS353 policy ticks · 11.77 simulated seconds · 45 policy calls · 220 bilateral contact ticks · 74.4 mm peak cube center height · direct reference: success
No video captured.
Measured object path
Cube center height (mm)
API control step latency (ms)
Cross embodiment evidence
YAM and SO101 live V3 transfer, plus separate scripted baselines
The trained V3 checkpoint was also evaluated by a live SO100 shadow policy FK/IK transfer controller through each target arm's APIs. It is not target camera conditioned. Open the dedicated YAM + SO101 report for actual model controlled outcomes, failures, videos, API calls and physical gates. The cards below are only scripted API baselines, excluded from policy success counts.
SO100 scripted API control
passed
7 completed commands · 101.7 mm maximum lift · 5.8 mm placement error
SO101 scripted API control
passed
7 completed commands · 96.0 mm maximum lift · 3.2 mm placement error
Reproducibility
What was run
| Checkpoint contract SHA-256 | 9bb09bc4fe31f86c72eef8f7e4821a8a176f25454b2bad164e1f9c4d1dfe97e5 |
|---|---|
| Checkpoint and processor bytes | verified |
| MuJoCo renderer | NVIDIA Corporation · NVIDIA A40/PCIe/SSE2 · 4.6.0 NVIDIA 595.91.07 |
| API regression suite | 68 tests; 0 failures; 0 errors |
| Policy input | Two 224×224 RGB cameras (front, wrist), six measured SO100 joint positions, exact task instruction |
| Policy output | 40 absolute six joint actions; first 8 consumed at 30 Hz |
| Dynamics and task | Original benchmark MujocoSO100Env and TaskTracker, 2 ms native physics timestep, gravity compensation |
| API boundary | Bearer authentication, lease epoch, sequence and freshness validation, named joint bounds, measured state, per tick journal |
| Seed set | [42000, 42001, 42002, 42003, 42004, 42005, 42006, 42007, 42008, 42009, 42010, 42011] |
The API returns two camera images and state atomically under its session lock. Each policy action passes through POST /v1/sessions/{id}/policy-step; the endpoint applies the original benchmark's 1/30 second control tick. Simulated time accumulates against MuJoCo's native 2 ms step. Direct and HTTP arms use the same scene and policy checkpoint.
Scope. This is a GPU backed MuJoCo software evaluation, not proof of physical hardware safety or WAN real time performance. The separate YAM/SO101 experiment uses live SO100 V3 inference and target arm API execution but does not feed target cameras into V3 or provide a target trained checkpoint. MuJoCo physics runs on CPU; NVIDIA EGL renders the videos.
Audit trail
Downloadable evidence
API run manifest · All API outcomes · Direct baseline outcomes · Three arm scripted results · Checkpoint and processor verification · NVIDIA renderer · Regression JUnit · Installed packages · GPU inventory · Policy server log · API server log · Artifact checksums
Publication record
This site contains the report pages and their linked evidence. Session lease values were removed from publication copies of API journals. The original evaluation files remain unchanged. Source and published file hashes record the exported files and redactions.