Observed arrival · 2026-11-05
Real2Sim: robot recordings, rebuilt by coding agents
A benchmark that tests whether coding agents can turn robot recordings into interactive simulations, then evaluates the reconstructions against visual and action-based evidence.
- For
- Robotics and coding-agent benchmark researchers
- Worth noticing
- The fixed π0.5-DROID policy ran on 57 scenes; three Argon Raw tasks instead received construction-fault zeros.
Field notes
The benchmark uses ten DROID episodes recorded on one Panda and camera rig; each provides three timestamped camera videos, stereo calibration, joint and gripper commands, and task text. Agents construct scenes in raw and RoboSimStudio conditions, with no human repairs before scoring. The report exposes task, model, and harness combinations; its fixed π0.5-DROID policy ran on 57 scenes, while three Argon Raw tasks received construction-fault zeros without a policy rollout.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue