Why a taught robot path cannot simply be sped up
A taught robot path cannot simply be sped up, because the robot controller decides how the arm accelerates and rounds each corner, and it decides differently at another speed. TOS Robotics measured this on FANUC robots in production, where one blended corner cut 14 mm when the arm started from a standstill and 24 mm when it arrived moving. We are therefore moving the timing out of the controller and into a planner of our own.
How the controller plans
On a teach pendant, the handheld control panel of an industrial robot, you store points with a speed, an acceleration setting and a termination type. FINE means stop on the point and settle. CNT 0 to 100 means blend into the next move without stopping, and a higher number rounds the corner more. The controller chooses how hard to accelerate and when to start a blend, and so which path the tool takes.
Our first learned controller, ACT, streamed a joint target every 4 ms, with the timing fixed at recording. From January 2026 the model predicts only the pick and place poses. The FANUC controller moves the arm in between, and since April it follows the belt with its own line tracking. That handed the path back to the controller.
What we measured
FINE costs time at each point. In FANUC’s simulator a 5-degree turn of the base joint at 100% speed took 300 ms and peaked at 13% of that joint’s top speed. That is because each move spent about 52 controller cycles, roughly 210 ms, speeding up and slowing down. In one day of line data the suction cup sealed 20 to 95 ms before the FINE stop was complete, and the robot still waited.
CNT is an overlap in time, not a distance. The faster the arm arrives, the earlier the blend starts and the more of the corner it cuts. We commissioned at 10% speed override, so clearances checked there had to be checked again at full speed.
We log the robot’s position every 4 ms and judge each change on tens of thousands of real cycles. A model of FANUC’s planner that we fitted from the outside overstated the gains 1.6 to 1.7 times on new data.
| Change | Expected | Measured |
|---|---|---|
| Approach height from 100 to 75 mm | 30 to 45 ms saved per pick | About 12 ms saved, and the cup touched the pepper while moving sideways about three times as often. Back to 100 mm. |
| Approach blend from CNT50 to CNT25 | A straighter path at contact | 37 to 38 ms slower per pick. Corner cut 15 to 6.5 mm, contact not straighter, no clear change in mispicks. Settled on CNT40. |
Owning the timing
In our cream dispensing development, a taught decoration must repeat with its own timing on a moving belt, which the controller’s planner and line tracking cannot do together. There our PC streams a joint target every 4 ms from our own jerk-limited planner. Its limits are lower than the controller’s, any alarm is a power-off stop, and the PC tracks the belt.
Bell peppers
In an offline estimate without belt tracking, our planner needed roughly half to nine tenths of the time the FANUC planner took on the line for the same joint moves. It gains by planning fewer stops and waits, not by moving harder. It has not run on a production robot. Our TOS Robot uses its own planner from the start.
Questions and answers
What is joint streaming?
Instead of sending points and letting the controller plan, the PC sends a joint target every few milliseconds, every 4 ms in our case. The PC then owns the timing, the limits, each deceleration and the belt tracking. Our planner uses the open source library Ruckig, which failed on about 1 in 1,200 of our updates until our patch, and on none of 28.8 million since.
Does TOS Robotics stream motion on the bell pepper line?
Not yet. The bell pepper line runs on programs in the FANUC controller with the controller's own line tracking. We use streaming in our cream dispensing development, and a streaming planner for bell peppers is in development.