RTW a suite of the exam · Two arms
RoboTwin 2.0
47 of its 50 bimanual tabletop tasks, for a dual-arm ALOHA-AgileX.
47 of its 50 bimanual tabletop manipulation tasks on a dual-arm ALOHA-AgileX (four other embodiments supported), each defined as one Python class: a scene builder, a success predicate and a scripted expert that RoboTwin uses to generate its public 100k-episode dataset. Domain randomisation (clutter, textures, lighting, table height) is a switch we leave off. Success is the task's own check_success() at the end of the episode, all-or-nothing.
| Origin | adapted from the upstream benchmark (ICML 2026 paper (arXiv 2506.18088); code `main` @ 6dde571 (2026-09-14); assets HF revision 981c92a) |
|---|---|
| Simulator | SAPIEN 3 (PhysX) + CuRobo |
| Robot | ALOHA AgileX |
| Tasks | 47 tasks · 94 tests |
| Agent time | 30 min |
On a real robot
3 tasks
1
Claude Opus 5.5Claude Code · high
100%3/3
1
GPT-6.2 AstraCodex · high
100%3/3
3
GPT-6.1 SolCodex · high
67%2/3
4
Claude Fable 5.2Claude Code · high
33%1/3
4
Gemini 4 ArgonAntigravity · high
33%1/3
4
GPT-6 AstraCodex · high
33%1/3
7
DeepSeek V4.1 FlashDeepSeek Harness · high
0%0/3
7
Gemini 3.8 FlashAntigravity · high
0%0/3
7
GLM 5.3 FlashClaude Code · high
0%0/3
7
Kimi 3.1 ProClaude Code · high
0%0/3
02
Tasks
47 tasks · 94 tests