KnowinGLOW: From EEF Actions to Continuous Motion
Observations and task context enter KnowinGLOW, which reasons, plans, and directly generates EEF actions. A motion interface sends these actions to the robot. New observations, robot state, and execution feedback return to the input to drive the next action.
KnowinGLOW ACTION TOKEN DECODING
Bridging Reasoning and Control with a Unified Action Representation
GLOW / MODEL
01 / CONTEXT
02/Reasoning & Planning & Action Generation
03 / EEF COMMAND
04 / EXECUTION
Observation & Task
Vision
Language
instructions
Robot state
KnowinGLOW
Spatial reasoning and task planning
Reach
Grasp
Transport
Place
Actions unfold as the task progresses
Autoregressive EEF
EEF → Motion Execution
EEF Command
EEF pose & gripper
Execute actions in sequence
Robot Execution
Actions in the
physical world
Reach · Grasp · Place
New observations · Robot state · Execution feedback
Unified Action Representation
Map spatial judgments to EEF pose
and gripper control
Autoregressive Generation
Generate subsequent actions
from new observations
Feedback Correction
Update subsequent actions from execution results
Pause
Replay
Observation & Task