Cosmos3 Nano Policy - Reactor

Cosmos3-Nano-Policy-DROID predicts an action chunk from a wrist camera, two exterior cameras, measured joint and gripper state, and a language instruction. Reactor hosts inference; your application executes the returned targets.

Variant and compatibility

Property	Value
Model identifier	`reactor/cosmos-nano-policy-droid`
Family	NVIDIA Cosmos3 Nano Policy
Checkpoint	Cosmos3-Nano-Policy-DROID
Reference embodiment	DROID: seven-joint Franka arm and gripper; RoboLab uses a Robotiq 2F-85 gripper
Action representation	Seven absolute joint positions in radians, plus normalized gripper closure
Prediction horizon	32 control steps; 8 values per step
Request mode	First chunk from observations; subsequent chunks require an executed-step acknowledgement
Documented model release	`0.4.2`; see release notes

The identifier selects the DROID variant. Other Cosmos checkpoints and robot embodiments do not inherit this observation/action contract. A seven-joint robot is not automatically compatible: joint ordering, gripper convention, camera geometry, and training distribution must match.

Path	What is provided
First actions	A Python example using synthetic camera fixtures; no simulator or local model weights
Simulation	An OpenPI-compatible gateway for NVIDIA RoboLab’s Cosmos3 client
Physical robot	Contract and integration guidance; these examples do not supply a validated hardware driver

Use the model

Start with Get your first actions, then choose simulation or robot integration. The reference specifies the wire contract. Reference clients live in the public simulator examples in Reactor’s public cookbook. Contact Reactor if you need hosted model access.

⌘I