laika-omni-c2-700
It learned to turn in place: 6 of 7 planned conditions met. The miss was a left-right tracking gap of 15 points against a limit of 10.
| Line | Walking in any direction |
| What changed | One new command, turning in place. The rewards stayed as they were, and gain randomization went back to 10%. |
| Checkpoint | Iteration 700. |
| In simulation | Passed some of its tests |
| On the real robot | Deployed |
| Exported | 2026-08-08 |
Lineage
Trained from laika-omni-s1e, resumed at checkpoint 500. laika-omni-s1e is not published on this hub.
Inputs and outputs
This section is read from the contract stamped inside the ONNX file. The robot
refuses a policy whose contract does not match its own. The same contract is in
io_contract.json.
The hub stamped this file for lucen/laika: its joints, actions, senses and rate are the robot's declared interface.
| Observation | 215 values: base angular velocity (3), gravity direction (3), velocity command (3), joint positions (12), joint velocities (12), previous action (12), gait phase (2), past frames (168) |
| Action | 12 values, one per joint, clipped to ±1 |
| Action scale | 0.5 on every joint |
| Control rate | 50 Hz |
| Gait cycle | 0.5 s |
| Starting base height | 0.38 m |
| Joint order | l_hip_pitch_joint, l_hip_roll_joint, l_hip_yaw_joint, l_knee_joint, l_ankle_pitch_joint, l_ankle_roll_joint, r_hip_pitch_joint, r_hip_roll_joint, r_hip_yaw_joint, r_knee_joint, r_ankle_pitch_joint, r_ankle_roll_joint |
Plant it was trained against
| Torque limit | RS00 11 N·m, RS02 17 N·m, RS06 12 N·m |
| Velocity limit | RS00 33 rad/s, RS02 20 rad/s, RS06 10 rad/s |
| Armature | RS00 0.0015, RS02 0.0032, RS06 0.007 |
| Joint friction | RS00 0.13, RS02 0.12, RS06 0.15 |
| Joint damping | 0.02 |
| Default stiffness and damping | RS00 12 and 0.6, RS02 20 and 1, RS06 30 and 1.5 |
| URDF and USD fingerprints | d6fab96bfc280a4d and 92b141c8dd200483 |
If the robot's motors differ from these, the simulator warns and the real robot refuses the policy.
Sim tests
Measured in simulation. 13 commands, 20 seeds each, at 80% motor power.
A command passes when every seed stays upright for the whole run and tracks 85% to 115% of the command. A check passes at 20 of 20 seeds.
| Test | Result | Reading |
|---|---|---|
| Stand still | Pass | 20 of 20 upright |
| Forward 0.15 m/s | Fail | 20 of 20 upright, 78% of the commanded speed |
| Forward 0.30 m/s | Pass | 20 of 20 upright, 90% of the commanded speed |
| Backward 0.10 m/s | Fail | 20 of 20 upright, 143% of the commanded speed |
| Backward 0.20 m/s | Pass | 20 of 20 upright, 103% of the commanded speed |
| Sideways left 0.10 m/s | Fail | 20 of 20 upright, 4% of the commanded sideways speed |
| Sideways right 0.10 m/s | Fail | 20 of 20 upright, -1% of the commanded sideways speed |
| Turn left 0.20 rad/s | Fail | 20 of 20 upright, 81% of the commanded turn rate |
| Turn right 0.20 rad/s | Fail | 20 of 20 upright, 59% of the commanded turn rate |
| Forward 0.20 m/s, turn left 0.20 rad/s | Pass | 20 of 20 upright, 85% of the commanded speed, 95% of the commanded turn rate |
| Forward 0.20 m/s, turn right 0.20 rad/s | Fail | 20 of 20 upright, 87% of the commanded speed, 29% of the commanded turn rate |
| Forward 0.15 m/s, sideways left 0.10 m/s | Fail | 20 of 20 upright, 78% of the commanded speed, -11% of the commanded sideways speed |
| Forward 0.15 m/s, sideways right 0.10 m/s | Fail | 20 of 20 upright, 78% of the commanded speed, 9% of the commanded sideways speed |
Real robot
The real-robot deployment policy, chosen after a comparison with laika-omni-c2-a800.
Note
It changed more than one thing. Three randomizations from another line stayed on, and its parent trained with them off.
Where it comes from
Lucen Robotics trained this policy for Laika in Isaac Lab. The inputs, outputs and plant come from the ONNX file's own stamp. The sim tests, where there are any, come from the training log's results.