X-WAM - Reactor

X-WAM predicts action chunks from three camera views, the current end-effector poses and gripper states, and a language instruction. Reactor’s endpoint returns actions and predicted robot states; it does not return generated video.

Variant and compatibility

Property	Value
Model identifier	`reactor/xwam`
Family	X-WAM
Checkpoint	`robotwin_sft`, RoboTwin fine-tuned
Reference embodiment	RoboTwin 2.0 `aloha-agilex`: two arms, two grippers, head and wrist cameras
Action representation	End-effector translation, axis-angle rotation, and gripper deltas for both arms
Prediction horizon	32 control steps; 14 values per step
Request mode	One outstanding request, one matching action reply
Documented model release	`0.3.2`; see release notes

The model identifier selects this hosted variant. It is not a checkpoint-selection parameter: the upstream pretrained and RoboCasa checkpoints are not interchangeable with this contract.

Path	What is provided
First actions	A runnable Python client using synthetic observations
Simulation	A gateway for the upstream RoboTwin 2.0 evaluation client
Physical robot	The observation/action contract and integration guidance; no validated hardware driver is supplied by these examples

The simulation configuration does not establish physical robot compatibility, including for hardware with a similar arm layout. Camera geometry, coordinate frames, gripper conventions, and the checkpoint’s training distribution still matter.

Use the model

Start with Get your first actions, then choose simulation or robot integration. The reference specifies every input and output field. The reference clients are in the public simulator examples in Reactor’s public cookbook. Contact Reactor if you need hosted model access.

⌘I