Skip to main content
Inspect Robots is an evaluation framework for robot policies. Use it with Dreamscale in one of two ways:
  • Record a dreamscale run as an eval log. Easiest if your robot is set up with Dreamscale.
  • Use --policy dreamscale inside your own Inspect Robots evaluation.

​
Record a run as an eval log

Add --inspect-robots to dreamscale run. The first run installs Inspect Robots.
The run works as usual. After each task, while the robot holds, grade it: y for success, n for failure, or a score from 0 to 1, optionally followed by a note. Then view the log:
Each instruction becomes a scene and each task a trial. For example, five timed trials of one task:
--inspect-robots works with every dreamscale run option except --hud.

​
Use Dreamscale as a policy

In the environment where you run Inspect Robots:
This installs Inspect Robots, signs you in if needed, and registers the dreamscale policy and the dreamscale_yam embodiment. Then select the policy with -P model=:
dreamscale_yam uses the rig you set up with dreamscale robots yam setup. To use your own YAM embodiment instead, run it at 30 Hz, for example --embodiment yam_arms -E control_hz=30. Supported models: molmoact2-bimanual-yam, pi05-yam, and perceptron-isaac05-yam (with -P checkpoint=ckpt_...).
Your embodiment sets the pace. If its step() returns faster or slower than 30 Hz, the model’s actions arrive at the wrong time and the robot holds more often, while the run still reports success. Pace the embodiment at 30 Hz.