JitterVLA: A Replay Benchmark for Motion Quality in VLA-Driven Human-Robot Manipulation

Faith Ogunwoye; Didar Seyidov; Konstantin Gubernatorov; Dzmitry Tsetserukou

cs.ROOct 9, 2026

JitterVLA: A Replay Benchmark for Motion Quality in VLA-Driven Human-Robot Manipulation

Authors: Faith Ogunwoye, Didar Seyidov, Konstantin Gubernatorov, Dzmitry Tsetserukou

Organizations: Intelligent Space Robotics Laboratory, Skolkovo Institute of Science and Technology Moscow, Russia

Abstract

Vision-language-action (VLA) policies issue sparse targets whose arrival at the robot controller is delayed and irregular, so a reasonable predicted action can still produce abrupt motion near people. Remedies are evaluated on different streams, delays and metrics. JitterVLA replays identical commands and arrival traces through alternative controllers, isolating the controller's effect. We apply it to synthetic reaches, hand trajectories from 20 people, three open VLA pipelines (838 episodes) and a physical UR3. Interpolation lowered modeled p99 jerk for every source, by 83% (615 to 106 m/s^3) on human motion with 200 ms delay, at a 2.0 cm tracking cost, and the effect replicated on 598 new episodes; jerk-limited generation lowered jerk further at a larger cost. Entropy-discontinuity associations had opposite, replicated signs in two pipelines, and entropy gating helped in one but not the other. On the UR3, interpolation lowered accelerometer-derived jerk by 3.0-14.0% in all six exploratory pairs, while a position-derived metric did not resolve the contrast. JitterVLA supports screening before human-centred studies; code and data will be released.

Figures & tables

Figure 1. JitterVLA. Commands from human motion and three VLA pipelines pass through inference and delivery timing; the same commands and arrivals are replayed through alternative controllers, and the resulting jerk–tracking trade-off identifies which conditions warrant physical and human-centred evaluation.

Source	N	Command form	Uncertainty
Synthetic reaches	10	5 Hz targets	–
Human, 3D-ARM-Gaze ( Lento et al., 2024 )	20 people	5 Hz targets	–
OpenVLA / BridgeData V2 ( Walke et al., 2023 )	80 / 200	per-step tokens	token entropy
pi0-FAST / DROID ( Khazatsky et al., 2024 )	80 / 199	10-step chunks	chunk entropy
pi0.5 / DROID	80 / 199	15-step flow chunks	none
UR3 (CB3 controller), 125 Hz	30 + 30 runs	replayed teleoperation motion	–

Table 1. Experimental design. Every source is replayed under all controllers with a paired delay realization. N is original/new-draw replication. 69,988 paired replays in total: Gamma delays

10{\times}7

(synthetic)

+\,20{\times}7{\times}10

(human)

+\,838{\times}61

(each VLA episode: zero delay once plus 6 means

\times

10 seeds)

=52{,}588

, and measured traces

3{\times}(10+10{\times}579)=17{,}400

(synthetic, human, OpenVLA and pi0-FAST ( Pertsch et al., 2025 ) units).

	Original		New draw
Source	$\Delta$ jerk	$\Delta$ RMSE	$\Delta$ jerk	$\Delta$ RMSE
Synthetic (10)	$-504$	$+2.39$	–	–
Human (20 people)	$-521$	$+2.32$	–	–
OpenVLA (80 / 200)	$-202$	$+0.78$	$-193$	$+0.77$
pi0-FAST (80 / 199)	$-371$	$+0.28$	$-343$	$+0.26$
pi0.5 (80 / 199)	$-374$	$+0.56$	$-363$	$+0.54$

Table 2. Interpolation minus stepped (B2

-

B0) at zero added delay: modeled p99 jerk (m/s 3 ) and tracking RMSE (cm), original and new-draw replication. Negative jerk favours interpolation; positive RMSE is a tracking cost.

Figure 2. Jerk–tracking frontier for hand trajectories from 20 people with 200 ms mean Gamma delay. Seeds are averaged within each person; 95% intervals resample people. Lower-left is better on both proxies; neither proxy measures comfort or task success. Four controller points: stepped B0 has the highest jerk and lowest error; interpolated B2 cuts jerk with slightly more error; jerk-limited B3 (25,m/s

^3

limit) cuts jerk further with more error; entropy-gated A2 is worse than B2 on both axes.

Figure 3. Entropy vs command discontinuity, original (filled) and new-draw (hollow) episodes kept separate; 95% unit-bootstrap intervals. The two panels use different action representations and discontinuity events, so their signs are not a model comparison. OpenVLA associations are negative in both original and new episodes; pi0-FAST associations are positive in both. No interval crosses zero.

Figure 4. Physical reality check. (a) The real UR3 CB3 setup; the translucent offset pose and its annotation are a conceptual illustration . (b) Simulator rehearsal of the replay pipeline (URSim, the Universal Robots offline simulator; software path only, not hardware). (c) Measured, exploratory v1.2a session: each marker is one paired direct-mode repeat’s B0 $-$ B2 difference in p99 tool-accelerometer-derived jerk. All six are positive; three repeats per delay are descriptive, not a significance test. Three panels: a UR3 photo with an illustrative translucent offset pose; a simulated jerk trace in which interpolated control has lower peaks than stepped control; six positive paired differences for no added delay and 200 ms mean delay.

Finding	Original	New draw	Held?
OpenVLA entropy vs goal step	$-.065$ [ $-.110$ , $-.021$ ]	$-.089$ [ $-.117$ , $-.061$ ]	yes
pi0-FAST entropy vs chunk jump	$.221$ [ $.169$ , $.271$ ]	$.217$ [ $.181$ , $.253$ ]	yes
B2 $-$ B0 jerk, 3 VLAs	$-202$ / $-371$ / $-374$	$-193$ / $-343$ / $-363$	yes
pi0-FAST controller contrasts	28 planned cells	28 / 28	yes
OpenVLA controller contrasts	28 planned cells	22 / 28	partly
pi0-FAST A2 vs best constant gain †	$-12.4$ (66/80)	$-5.6$ (162/199)	direction

Table 3. What replicated on new, disjoint episode draws (planned v1.4 test unless noted). Intervals are 95% unit bootstraps.

Robotic Control

Robotic control studies how machines use sensing and computation to produce reliable movement and interaction in changing environments.

See all 584 papers

GROOVE: Geometry-Guided Reduction of Operational-Space Jerk in VLA Execution
Sep 12, 2026Sangho Yun, Minsoo Kim, Minwoo Cho +1Robot Motion Generation Trajectory Optimization
Department of Computer Science and Engineering, POSTECH, Pohang, South Korea. · Graduate School of Artificial Intelligence, POSTECH, Pohang, South Korea.
MotionVLA: Injecting Geometric Motion into Vision-Language-Action Model
Jun 6, 2026Shanglin Yuan, Weiheng Zhao, Xianda Guo +4Long-Horizon Robotic Manipulation Visuomotor Policy Learning
Huazhong University of Science and Technology · D-Robotics · Wuhan University
SkipVLA: Skipping VLA Steps with Classical Planning for Fast Robot Manipulation
Sep 17, 2026Kaivalya Agrawal, Md Ashiqur Rahman, Raymond A. Yeh +1Long-Horizon Robotic Manipulation Efficient VLA Model Inference
Department of Computer Science, Purdue University

Sep 12, 2026J/K move · Enter open · S save

GROOVE: Geometry-Guided Reduction of Operational-Space Jerk in VLA Execution

Sangho Yun, Minsoo Kim, Minwoo Cho, Hwanjo Yu

Department of Computer Science and Engineering, POSTECH, Pohang, South Korea. · Graduate School of Artificial Intelligence, POSTECH, Pohang, South Korea.

Robot Motion Generation Trajectory Optimization VLMs for Robotics

Open paper AskPDF arXiv

Chunked vision language action (VLA) policies execute several commands per query, but jerk within chunks and across replanning boundaries can induce oscillatory motion and sharp actuator transients. We present GROOVE, an online regulator that searches directional correction regions around the raw three dimensional end effector (EEF) path, without retraining or additional VLA inference. It optimizes the new chunk using delivered commands as boundary conditions, reducing boundary and within chunk jerk while bounding cumulative translation and local axis angle deviation from the raw plan after every command. Using quadratic programs (QPs), GROOVE generates a cube reference and thirteen directional candidates, then selects the one with the lowest command space jerk under a reference relative deviation cap. On a held out LIBERO benchmark, GROOVE achieves the largest reductions among the evaluated methods, reducing translational and rotational EEF jerk by 33.02% and 43.42%, respectively, with task success of 95.75% versus 93.75% for raw execution. Across 50 matched UR5e pairs with measured execution timing, it reduces translational and rotational tool center point (TCP) jerk by 16.39% and 19.49% and joint current slew by 29.09%.

Abstract

Robotic Control

Explore similar work

GROOVE: Geometry-Guided Reduction of Operational-Space Jerk in VLA Execution

MotionVLA: Injecting Geometric Motion into Vision-Language-Action Model

SkipVLA: Skipping VLA Steps with Classical Planning for Fast Robot Manipulation