Benchmarks
Standard tricks, honest numbers
A fixed set of tasks per robot, trained on kualia the same way your task is — same checks, same PPO defaults, same CPU servers — and evaluated on up to 30 deterministic episodes. Every success rate comes with a 95% interval, so a lucky run looks like one.
Suite v0 (suite defined, runs pending) · 0 of 10 tasks run
| Robot | Trick | Success means | Budget | Success · 95% CI | Train time |
|---|---|---|---|---|---|
| Dynamixel 2R | Reach a target point and hold | Tip within 2 cm of the target at the end of the episode | 50K steps | not run yet | — |
| SO-101 arm | Reach a marker on the table | Gripper within 2 cm of the marker at the end | 300K steps | not run yet | — |
| SO-101 arm | Press the button and hold still | Button pressed at the end of the episode | 300K steps | not run yet | — |
| Low-cost arm | Push a block onto the target | Block centre inside the target | 200K steps | not run yet | — |
| Crazyflie 2 | Take off and hover at 0.5 m | Within 5 cm of 0.5 m for the last second | 300K steps | not run yet | — |
| Open Duck Mini | Walk forward 1 m | Base 1 m ahead, still standing | 2M steps | not run yet | — |
| ROBOTIS OP3 | Walk forward 1 m | Base 1 m ahead, still standing | 5M steps | not run yet | — |
| Unitree Go2 | Trot forward 2 m | Base 2 m ahead, upright | 3M steps | not run yet | — |
| Inverted pendulum | Balance the pole | Pole up for the whole episode | 30K steps | not run yet | — |
| Hopper | Hop forward 2 m | 2 m along x without falling | 300K steps | not run yet | — |
How each number is produced
- The task is written by alia from the brief in the table and must pass all 13 checks before training.
- PPO (Stable-Baselines3 defaults) on kualia's CPU servers, with the budget shown; no hand-tuning per task.
- Evaluation: the final policy, deterministic, on up to 30 fixed seeds (0–29).
- Success rate with a 95% Wilson interval; the mean return and training time are reported alongside.
- Each row links to the run once published, so the task code, curve and replays can be inspected.
Run it yourself
Every benchmark is an ordinary kualia task. Pick the robot, paste the trick and success criterion, and train with the same budget — your result page shows the same interval.
Open the Studio