single mention · model architecture
Predicting end-effector trajectory traces as intermediate visual plans
A Visual Reasoning Trace is a predicted intermediate plan: the model generates the 2D polyline of the robot end-effector's future positions on the image alongside the next action command.1 It is defined as a polyline of points subsampled from the future episode horizon, with the first point the current end-effector location, and is generated before the action tokens to ground them spatially.2
Trade-off Traces give precise, editable test-time steering that avoids the ambiguity of language prompts.3
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
we generate the 2D trajectory of the end-effector and train the model to predict this trajectory alongside the next action command
— MolmoAct Action Reasoning Models that can Reason in Spacepoints are subsampled evenly from the future horizon of the episode between the current frame and the terminal frame
— MolmoAct Action Reasoning Models that can Reason in Spacevisual reasoning traces are unambiguous, easily edited, and generalize across tasks without large text–action corpora or brittle language patterns
— MolmoAct Action Reasoning Models that can Reason in Space
Also called Visual Reasoning Trace.
Filed under model architecture › prediction head