YAM and SO101: pick and place through Robot APIs

YAM: model controlled API transfer

Physical success. The SO100 V3 checkpoint observed a live SO100 shadow scene; each newly inferred model action was converted by forward kinematics, a two anchor workspace transform, and target arm inverse kinematics before an authenticated target API command. This is not native YAM camera conditioned policy inference, retraining, or a replayed/scripted path.

Controller corrections were chosen on seed 42001; seeds 42002 and 42003 are held out follow ups using the fixed final controller: 2/2 physical successes for YAM. This small, selected seed set is not a fleet level reliability estimate. The attempt table distinguishes tuning history from follow up and never replaces a failure with a later success.

Every attempt, including failures

Actual API call sequence

Counts for the selected run: 337 target observations, 337 setpoints, 350 physics steps, 0 IK failures. Example tick from the actual trace (no authentication secrets):

{
  "tick": 0,
  "api_observation_sequence": 0,
  "api_command_sequence": 0,
  "api_accepted_sequence": 0,
  "source_target": [
    0.5309028029441833,
    -1.3722734451293945,
    1.3263784646987915,
    1.6236151456832886,
    -2.1009674072265625,
    0.699999988079071
  ],
  "target_tool_goal_m": [
    0.24854906658195788,
    -0.0016476439025129817,
    0.15955210704380535
  ],
  "target_joint_target": {
    "joint1": -0.004498660669756926,
    "joint2": 1.4983425563202757,
    "joint3": 1.401610764931806,
    "joint4": -1.3009176420426312,
    "joint5": 0.0029823384967683207,
    "joint6": -0.0007929796696357223,
    "gripper": 0.03752399936097009
  },
  "target_positions": {
    "joint1": -0.0009319544215701186,
    "joint2": 1.4116157395478968,
    "joint3": 1.4593794952215824,
    "joint4": -1.2999528997217482,
    "joint5": 0.0003000976727029928,
    "joint6": -0.00015531627067669914,
    "gripper": 0.037524161295854064
  },
  "target_object": {
    "position_m": [
      0.2500009098,
      -2.715396712e-17,
      0.10006832832704898
    ],
    "quaternion_wxyz": [
      0.7071067811865477,
      0.0,
      0.0,
      0.7071067811865477
    ],
    "contacts": [
      {
        "geom": "pickup_pedestal",
        "body": "world",
        "normal_force_n": 0.061353299067769144
      },
      {
        "geom": "pickup_pedestal",
        "body": "world",
        "normal_force_n": 0.061353299067769144
      },
      {
        "geom": "pickup_pedestal",
        "body": "world",
        "normal_force_n": 0.061353299067769144
      },
      {
        "geom": "pickup_pedestal",
        "body": "world",
        "normal_force_n": 0.061353299067769144
      }
    ],
    "left_finger_force_n": 0,
    "right_finger_force_n": 0,
    "bilateral_contact": false,
    "supported": true,
    "lift_m": -5.304070083347878e-05
  }
}

Physical measurements

Acceptance requires at least 2 simulated seconds unsupported bilateral lift, one final second released and settled on the drop pedestal, less than 40 mm XY destination error, and less than 1 mm final one second drift. Selected run result: 1.04 s settled; 0.00 mm drift. All values are MuJoCo contacts and positions, not video interpretation.

Measured object path

Object lift (mm)

Bilateral contact

Outcome and mapping · Full per tick model/API/physics trace · API episode journal · HTTP server log · Model load log

SO101: model controlled API transfer

Physical success. The SO100 V3 checkpoint observed a live SO100 shadow scene; each newly inferred model action was converted by forward kinematics, a two anchor workspace transform, and target arm inverse kinematics before an authenticated target API command. This is not native SO101 camera conditioned policy inference, retraining, or a replayed/scripted path.

Controller corrections were chosen on seed 42001; seeds 42002 and 42003 are held out follow ups using the fixed final controller: 2/2 physical successes for SO101. This small, selected seed set is not a fleet level reliability estimate. The attempt table distinguishes tuning history from follow up and never replaces a failure with a later success.

Every attempt, including failures

Actual API call sequence

Counts for the selected run: 338 target observations, 338 setpoints, 351 physics steps, 0 IK failures. Example tick from the actual trace (no authentication secrets):

{
  "tick": 0,
  "api_observation_sequence": 0,
  "api_command_sequence": 0,
  "api_accepted_sequence": 0,
  "source_target": [
    0.5309028029441833,
    -1.3722734451293945,
    1.3263784646987915,
    1.6236151456832886,
    -2.1009674072265625,
    0.699999988079071
  ],
  "target_tool_goal_m": [
    0.18588802955942904,
    6.789120500600258e-05,
    0.194555860062779
  ],
  "target_joint_target": {
    "shoulder_pan": -0.0016564132629382545,
    "shoulder_lift": -0.9555073943692493,
    "elbow_flex": 0.1453621677642259,
    "wrist_flex": 1.6333634554968275,
    "wrist_roll": 8.610624509433877e-19,
    "gripper": 0.9999999799978735
  },
  "target_positions": {
    "shoulder_pan": -0.0016388433693159225,
    "shoulder_lift": -0.9543858932695792,
    "elbow_flex": 0.14676718339919942,
    "wrist_flex": 1.633470392622948,
    "wrist_roll": -6.951533982584702e-08,
    "gripper": 0.9999963786772659
  },
  "target_object": {
    "position_m": [
      0.18460905,
      5.799917456000146e-06,
      0.10007208234258336
    ],
    "quaternion_wxyz": [
      0.9997186370071134,
      6.2436799921968536e-18,
      1.8016201939805703e-18,
      -0.023720177500168776
    ],
    "contacts": [
      {
        "geom": "pickup_pedestal",
        "body": "world",
        "normal_force_n": 0.06135329905581372
      },
      {
        "geom": "pickup_pedestal",
        "body": "world",
        "normal_force_n": 0.06135329905581372
      },
      {
        "geom": "pickup_pedestal",
        "body": "world",
        "normal_force_n": 0.061353299055813715
      },
      {
        "geom": "pickup_pedestal",
        "body": "world",
        "normal_force_n": 0.061353299055813736
      }
    ],
    "left_finger_force_n": 0,
    "right_finger_force_n": 0,
    "bilateral_contact": false,
    "supported": true,
    "lift_m": -5.3040697683498506e-05
  }
}

Physical measurements

Acceptance requires at least 2 simulated seconds unsupported bilateral lift, one final second released and settled on the drop pedestal, less than 40 mm XY destination error, and less than 1 mm final one second drift. Selected run result: 1.04 s settled; 0.00 mm drift. All values are MuJoCo contacts and positions, not video interpretation.

Measured object path

Object lift (mm)

Bilateral contact

Outcome and mapping · Full per tick model/API/physics trace · API episode journal · HTTP server log · Model load log

Separate scripted API baselines. not policy evidence

These are precomputed API commands, retained only as physical engine and API baselines. Their success is excluded from all model controlled claims.

Authenticated API · YAM

YAM: pick, carry, place, release

PASSEDPhysics success gate

900 / 25Distinct frames / video FPS

36.0 sContinuous MuJoCo video

7Completed API commands

99.9 mmMaximum unsupported lift

1.02 mmFinal placement error

Observed motion phases: Ready → Approach → Grasp → Lift → Transfer → Lower → Release → Retreat. The object moved through genuine frictional MuJoCo contacts; the evaluator did not weld it to the gripper or overwrite its pose.

Visual timeline

Physical task measurements

Unsupported bilateral lift	10.28 simulated seconds
Final placement error	0.001023 m
Final one second object drift	0.000000 m
Settled release	25 frames
Object mass	0.025 kg
Object size	[0.016, 0.024, 0.04] m
Pickup position	[0.2500009098366757, -2.7153967119061406e-17, 0.10012136902788246] m
Destination position	[0.17417713860911788, 0.17933949766865334, 0.10012136902788246] m
Contact model	MuJoCo frictional contacts; no weld or object pose writes

Measured object XY path

Object lift, millimeters

Bilateral jaw contact, 0 or 1

Result JSON · 900 frame state trace · API episode journal · Command runs · API server log

Authenticated API · SO101

SO101: pick, carry, place, release

PASSEDPhysics success gate

900 / 25Distinct frames / video FPS

36.0 sContinuous MuJoCo video

7Completed API commands

96.0 mmMaximum unsupported lift

3.23 mmFinal placement error

Observed motion phases: Ready → Approach → Grasp → Lift → Transfer → Lower → Release → Retreat. The object moved through genuine frictional MuJoCo contacts; the evaluator did not weld it to the gripper or overwrite its pose.

Visual timeline

Physical task measurements

Unsupported bilateral lift	10.20 simulated seconds
Final placement error	0.003233 m
Final one second object drift	0.000000 m
Settled release	25 frames
Object mass	0.025 kg
Object size	[0.016, 0.024, 0.04] m
Pickup position	[0.18460904999305738, 5.799917456049495e-06, 0.10012512304026686] m
Destination position	[0.14039968358821892, -0.10456627691997465, 0.10012512304026686] m
Contact model	MuJoCo frictional contacts; no weld or object pose writes

Measured object XY path

Object lift, millimeters

Bilateral jaw contact, 0 or 1

Result JSON · 900 frame state trace · API episode journal · Command runs · API server log

Evaluation and provenance

Each arm used its calibrated pick_place scene with a 25 g object. API command completion was necessary but not sufficient: the gate checked physical lift, release near the target, and settled end state. MuJoCo physics ran on CPU; NVIDIA A40 EGL rendered the videos.

Scripted baseline data · Offline 52 pose FK/IK feasibility probe · NVIDIA renderer · Downloaded evidence 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.