← All RL teachers

One teacher • motion experiment

Lift-Cube: speed & smoothness

Compare the original teacher with the motion trial on the same three starting conditions. Every video plays at 1× real time, 50 fps. These episodes were selected before evaluating the trial.

Motion targets not met — further Lift tuning needed
Trial success: 25/128 validation · Not qualified / not run independent confirmation.
The original teacher passed 118/128 and 117/128. Other teachers remain paused pending your review.

Motion across all 128 evaluation episodes

Values are the 95th percentile of episode measurements, including failures. Acceleration is estimated from recorded 20 ms velocity samples. A target step is the change in commanded joint angle, not physical joint travel.

MeasurementOriginalMotion trial
Peak joint speed8.07 rad/s7.70 rad/s
Acceleration estimate58.92 rad/s²58.55 rad/s²
Largest target step4.56 rad4.61 rad

Episode 1

Original teacher Task succeeded

Motion trial Task failed

Episode 2

Original teacher Task succeeded

Motion trial Task succeeded

Episode 3

Original teacher Task succeeded

Motion trial Task succeeded

The physical task, success rules, 20-second episode limit, observations and action interface are unchanged. The trial changes only training rewards for motion. The original checkpoint is retained.