LingBot-VA LIBERO

LingBot-VA LIBERO takes an external camera, a wrist camera, and a language task. It returns a chunk of end-effector delta commands and uses the executed actions and observed video to predict the next chunk. It has no proprioception input or generated-video output on this endpoint.

Property	Value
Model identifier	`reactor/lingbot-va`
Upstream checkpoint	LingBot-VA posttrain LIBERO-Long
Reference embodiment	LIBERO’s Panda arm, gripper, and `OSC_POSE` controller
Input tracks	`agentview`, `eye_in_hand`; RGB, 128 × 128 reference resolution
Output	`(16, 7)` action chunk; first episode chunk has four conditioning rows to skip
Reference control rate	20 Hz: 0.6 seconds for the first 12 executed rows, 0.8 seconds for subsequent 16-row chunks
Documented model release	`0.2.3`; see release notes

The control rate describes execution, not inference throughput. Replies arrive after inference and transport delay; the reference simulator pauses between chunks.

Path	What is provided
First actions	Synthetic camera fixtures and a Python client; no local model weights
Simulation	A LIBERO/robosuite environment and Reactor bridge
Physical robot	Contract and adapter guidance; no validated hardware driver, IK implementation, or calibrated mapping

Follow Get your first actions, then simulation or robot integration. The reference defines the wire and action semantics.

Get started

Models

LingBot-VA LIBERO