MJP  a suite of the exam · Quadruped, two arms and one arm

MuJoCo Playground

Quadrupeds that walk, get up and balance, and manipulation for ALOHA 2 and a Franka Panda.

Thirteen environments of Google DeepMind's MuJoCo Playground, run in plain MuJoCo on the CPU. Nine are quadruped locomotion: the Unitree Go1 tracks a joystick on flat and rough ground, gets up from a fall and balances in a handstand and a footstand; Boston Dynamics Spot tracks a joystick, with and without a commanded gait, and gets up; Google's Barkour vB tracks a joystick while it is pushed. Four are manipulation: ALOHA 2 hand-over and peg insertion, and a Franka Panda that opens a cabinet and places a cube in a given orientation. Playground trains with dense rewards, so each task has a success test of ours, built from its termination rules and reward targets.

Originadapted from the upstream benchmark (MuJoCo Playground (arXiv 2502.08844); code `main` @ 4057c14)
SimulatorMuJoCo 3.3.7 (plain MuJoCo on the CPU; upstream trains in MJX / MuJoCo Warp)
RobotALOHA 2, Boston Dynamics Spot, Franka Panda, Google Barkour vB, Unitree Go1
Tasks13 tasks · 27 tests
Agent time30 min
01

Results

7 tasks
1 GPT-6.2 AstraCodex · high 100%7/7
2 GPT-6 AstraCodex · high 57%4/7
3 Claude Fable 5.2Claude Code · high 43%3/7
3 GPT-6.1 SolCodex · high 43%3/7
5 Gemini 4 ArgonAntigravity · high 29%2/7
6 Claude Opus 5.5Claude Code · high 0%0/7
6 DeepSeek V4.1 FlashDeepSeek Harness · high 0%0/7
6 Gemini 3.8 FlashAntigravity · high 0%0/7
6 GLM 5.3 FlashClaude Code · high 0%0/7
6 Kimi 3.1 ProClaude Code · high 0%0/7
02

Tasks

13 tasks · 27 tests