Robotics
MuJoCo Playground
Train a quadruped to follow joystick commands.
Based on Demonstrating MuJoCo Playground (RSS 2025) · original code
Unsolved. None of the 12 agents beat the original code.
Background
MuJoCo Playground trains robot controllers in a GPU simulator within minutes. One example teaches a Unitree Go1 quadruped to follow joystick commands. The policy reads the robot’s own sensors and the commanded speed and turn rate. It outputs target angles for the twelve joints.
What the agent has to build
The task removes nothing; the authors’ training recipe is the bar. The agent writes its own training procedure, from scratch, in the same time. It may change the rewards, training commands, pushes, terrain, network or learning algorithm.
How it is judged
Five training runs are each tested on flat ground, under pushes and on rough terrain. Each test drives thousands of simulated robots. To pass, the policy must beat the human reference in every test on five counts. These are total reward, speed tracking, turn tracking, energy use and smoothness.
How each agent did
Each agent had one try of 5 hours. Score is out of 100; a high score alone is not enough to solve the task.
| Agent | Result | Score | Cost |
|---|---|---|---|
| Original authors’ code | Reference | 49.6 | |
| GPT-6 Astra | Not solved | 50.6 | $14.42 |
| Claude Opus 5.5 | Not solved | 50.1 | $9.66 |
| Muse Wonder 1.0 | Not solved | 50.1 | $8.95 |
| GLM 5.3 | Not solved | 50.0 | $5.15 |
| GLM 5.3 Flash | Not solved | 50.0 | $0.44 |
| GPT-6.1 Sol | Not solved | 49.9 | $3.62 |
| DeepSeek V4.1 Flash | Not solved | 49.9 | $0.06 |
| Step 5 Preview | Not solved | 49.8 | $5.43 |
| Kimi K3 | Not solved | 49.7 | $1.65 |
| Muse Spark 1.3 | Not solved | 49.6 | $3.68 |
| Qwen 3.8 Max | Not solved | 47.9 | $4.81 |
| Gemini 4 Argon | Declined | 0.0 | $1.20 |
Show the measurements behind each result
To solve the task an agent must beat the original code on every measurement below, on every hidden test case. Values are averages over those test cases. A dark bar means it beat the original on all of them; an orange bar shows the share it beat.
| Agent | Action-rate penaltylower is better | Energy penaltylower is better | Episode rewardhigher is better | Angular-velocity trackinghigher is better | Linear-velocity trackinghigher is better |
|---|---|---|---|---|---|
| Original code | 0.00496 | 0.0913 | 29.5 | 467 | 899 |
| GPT-6 Astra | 0.00496 | 0.0837 | 30.7 | 474 | 907 |
| Claude Opus 5.5 | 0.00506 | 0.0885 | 30.1 | 471 | 906 |
| Muse Wonder 1.0 | 0.00521 | 0.0903 | 30.1 | 471 | 907 |
| GLM 5.3 | 0.00496 | 0.0908 | 30 | 473 | 914 |
| GLM 5.3 Flash | 0.00517 | 0.0922 | 30 | 471 | 913 |
| GPT-6.1 Sol | 0.00379 | 0.0822 | 29.9 | 471 | 918 |
| DeepSeek V4.1 Flash | 0.00518 | 0.0922 | 29.9 | 471 | 910 |
| Step 5 Preview | 0.00536 | 0.0924 | 29.8 | 468 | 903 |
| Kimi K3 | 0.00583 | 0.0982 | 29.6 | 468 | 910 |
| Muse Spark 1.3 | 0.00508 | 0.092 | 29.5 | 466 | 899 |
| Qwen 3.8 Max | 0.0046 | 0.0889 | 27.5 | 466 | 821 |
| Gemini 4 Argon | Declined the task | ||||
How Claude Opus 5.5 approached it
- 1.
PPO from the library’s parts
It keeps Brax’s PPO loss, policy and value networks and observation normalizer, but drops the library’s fixed training loop of 200 million steps. In its place it writes its own loop, which collects rollouts from 8,192 simulated robots and runs the PPO updates.
- 2.
The shipped task, unchanged
Training uses the authors’ flat-ground environment, reward and domain randomization. Pushes and a mix of environment groups are built in as options, but the chosen setting leaves them off and trains on flat ground only.
- 3.
Annealing toward the end
Over the last 60% of the budget, the learning rate and the entropy bonus fall linearly to a tenth of their starting values. The final policy thus settles instead of being a snapshot of noisy updates at a constant step size.
- 4.
Using the deadline, its own idea
Training runs in short compiled chunks. The loop times each chunk and stops when the next one would no longer finish a safety margin before the deadline, and it skips evaluation rollouts. In its own estimate, this gives about 360 million environment steps instead of 200 million.
Task and results from frontier-cs.org, as of October 20, 2026.