SWIRL-Lab2026cc-by-4.0
BiGym 2.0 Coding-Agent Rollouts
Evaluation episodes from coding-agent results in BiGym 2.0, containing 5,400 episodes across 54 frozen programs from two LLM agents (Claude Opus, GPT-6 Astra) solving nine humanoid loco-manipulation tasks. Each episode includes full simulator state, actions, rewards, and code execution traces.
Downloads22
Episodes5400
Why This Matters for Physical AI
This dataset provides a comprehensive benchmark of LLM-generated policies for humanoid manipulation, enabling evaluation of coding-agent-based robot learning and policy development approaches.
Technical Profile
- Modalities
- proprioceptionsimulator_state
- Robot Embodiments
- humanoid
- Action Space
- joint_positions
- Environment
- simulation
- Task Types
- loco-manipulationmanipulation
- Episodes
- 5400
- Data Format
- npz
- Annotation Types
- action_labelsreward_labelscode_traces
- License
- cc-by-4.0
Access
Need custom proprioception data?
Claru builds purpose-built datasets for simulation applications with dense human annotations and quality assurance.
Request a Sample Pack