Motion smoothness metrics

Motion smoothness metrics are scores that measure how smoothly a movement unfolds, computed from a speed or acceleration signal. Common ones are jerk, the rate of change of acceleration, log dimensionless jerk (LDLJ), which normalizes jerk for duration and amplitude, and spectral arc length (SPARC), which measures how spread out a movement's frequency content is. In robot learning, they help rank demonstrations and rollouts for review, but they measure the motion, not whether the task succeeded.

What are motion smoothness metrics?

A smooth movement speeds up, peaks, and slows down once. A rough one hesitates, overshoots, and corrects. Motion smoothness metrics turn that difference into a number. They come from motor neuroscience and rehabilitation research, where they track how arm movement recovers, and robot learning teams now use them to score episodes so they can sort a large set of recordings by how rough each one looks.
Every metric starts from a motion signal, usually the speed of a joint or the end-effector, taken from the robot's proprioception or an inertial measurement unit.

Key takeaways

  • Jerk, LDLJ, and SPARC each score a different aspect of how smoothly something moved.
  • LDLJ and SPARC are dimensionless, so you can compare movements of different lengths and sizes.
  • Smoothness measures the motion, not the outcome, so a smooth episode can still fail its task.

How it works

  • Jerk is the third derivative of position, or how fast acceleration changes. Jerk RMS reports it in physical units, which helps find a local rough spot but makes movements of different sizes hard to compare.
  • Log dimensionless jerk (LDLJ) integrates squared jerk over the movement, scales it by the movement's duration and peak speed so the result has no units, then takes the logarithm. Values are negative, and more negative means less smooth.
  • Spectral arc length (SPARC) takes the Fourier magnitude spectrum of the speed profile, normalizes it, and measures the length of that curve up to an adaptive cutoff frequency. A single smooth movement keeps its energy at low frequencies and gives a short arc. Corrections and tremor add higher-frequency content and lengthen it. Values are negative, and more negative means less smooth.
SPARC was introduced in 2015 by Balasubramanian and colleagues as a refinement of their earlier spectral arc length metric from 2012, and the same paper analyzes LDLJ alongside it. Tools such as the open source demo_quality_scorer plugin for FiftyOne compute these metrics in sliding windows over an episode, summarize each episode with the median window and the worst window, and then compare each episode with the rest of the batch.
A smooth speed profile scores well and can still fail the task.

Plot the speed trace, then check what the rough spots really were

A smoothness score tells you where to look, not what to keep. Watch for:
  • rough speed traces that are really a regrasp, not a bad motion
  • rough scores caused by a sensor dropout, not by the motion
  • short episodes, where a sliding window has too few samples to score reliably
In Voxel51, DROID holds 1,000 real-world manipulation episodes with end-effector speed signals, motion stats such as peak end-effector speed, and grasp counts.

Robotics Droid

1,000 episodes

Real-world robot manipulation episodes from DROID, stored as MCAP with three synced cameras and fields for task prompt, success, and number of grasps.

Try this: Filter num_grasps to 3 or more, open an episode, and find the stops and restarts in its end-effector speed signal.

Explore DROID in Voxel51

Free with a work or .edu email.

Go deeper

Why it matters

Nobody can watch every demonstration in a large dataset, so smoothness scores help decide which ones to watch first. They surface hesitations, regrasps, and jittery corrections that can teach a policy bad habits. They are triage, not a filter. When Voxel51 scored 127 simulated rollouts of one RoboLab task, an episode that ranked 8th-smoothest and cleared every outlier check had still failed the task, as described in Motion Smoothness Doesn't Predict Robot Task Success. Pair smoothness with a task success label or success detection before you drop or keep anything.
Scores can also be relative. A scorer that fits them against the batch gives the same episode a different score in a different batch.
Smoothness ranks episodes for review. It does not say which ones succeeded.

Frequently asked questions

What is log dimensionless jerk?

Log dimensionless jerk (LDLJ) is a smoothness score computed by integrating squared jerk over a movement, scaling it by duration and peak speed to remove units, and taking the logarithm. Values are negative, and values closer to zero mean smoother motion.

What is spectral arc length (SPARC)?

SPARC measures the length of a movement's normalized speed spectrum up to an adaptive cutoff frequency. Smooth movements concentrate their energy at low frequencies and produce values closer to zero.

Does a smoother robot episode mean a better demonstration?

Not on its own. Smoothness says nothing about whether the task was completed, so a smooth episode can still put the object in the wrong place. Use smoothness to prioritize review, and check task success separately.
black and white photo of Jesse Mostipak
Director of Growth
Jesse Mostipak is the Director of Growth at Voxel51, where the work is helping humans find and trust what the brand knows, and teaching the Google knowledge graph and the LLMs answering on their behalf to do the same. That question, how knowledge gets built inside a system, is one Jesse has been chasing for years. Earlier versions of it ran through a New York City high school science classroom, data science and machine learning, and developer relations at Kaggle, Posit (formerly RStudio), and Baseten. The answer doesn't change much depending on whether the learner is a teenager, a software engineer, or a knowledge graph. Jesse holds a Master's in Education from CUNY Hunter College. LinkedIn
See all articles by Jesse Mostipak
Last updated November 19, 2026

Building visual or physical AI?

Let's talk.