FAR: Faster Agentic Robotics
with Fewer Tokens

Mayank Chandak*Shi Chen*Yu-Lun Chou* Vamsi Grandhi*Mayanka Medhe*Jeffrey Ichnowski

Robotics Institute, Carnegie Mellon University

*Equal contribution; listed alphabetically by surname.

Full paper on arXiv and code on GitHub coming soon.

Method

FAR method: retain useful code from earlier tasks in a skill library, adapt it to the new task with a code agent, then execute and refine on the robot or simulator. Examples: reusing code in simulation (set up mug to apple to sink), reusing code to stack more cups (2 to 3 cups, 3 to 5 or 6 cups, a six-cup pyramid), and reusing code for hanging (hang a cup to hang a cap).
FAR retains code from a completed donor task (a previously solved task whose code it reuses), adapts the task logic and calibration for a recipient, and refines the program using execution feedback. The lower panels show simulation and Kinova examples.

Rebuilding perception and motion routines for each task consumes time and model tokens. FAR lets an agent reuse tested code from a similar task on the same arm.

Simulation Trials

Five Embodied First Exam tasks. The transfer agent receives code from a solved mug-placement task (CoffeeSetupMug); the baseline receives no donor library.

Agent time is the summed duration of coding-agent turns, excluding robot execution and operator waits. Three seeds per condition, Opus 5.5 at high effort.

Videos play at 2×. Click any video to enlarge it.

Physical Trials

Ten transfers on two Kinova Gen3 arms. The baseline receives a demonstration, calibration, and the robot interface; the transfer agent also receives the donor skill library.

Agent time is the summed duration of coding-agent turns, excluding robot execution and operator waits. Values average two seeds, Opus 5.5 at medium effort.

Videos play at 4×. Click any video to enlarge it.

Emergent System Behaviors

Making space first

The robot makes space by first moving a cup out of the way.

8×

Placing on the fingertip

The robot learns to place the cup upright on the fingertip instead of along the finger.

4×