# Does Personalized Motion Scaling Actually Fix Teleoperation Delay?

Communication delay is the core unsolved problem in telerobotics — and new research from Sai Jiang and Zonghe Chua suggests that personalized, context-aware motion scaling can deliver up to 20–25% performance gains in key metrics compared to unassisted operation. That number comes from a 20-participant study where subjects performed delayed reaching tasks in a virtual simulator, with evaluation extended to a telesurgical robot. The performance improvement was most pronounced at longer delays — exactly the regime that matters for humanoid teleoperation over real-world networks.

The study's core finding cuts against a dominant assumption in deployed teleop systems: that uniform motion scaling gains are good enough. They aren't. The researchers show that scaling gains need to vary by participant, movement direction, distance, and delay magnitude to meaningfully close the precision gap. Critically, personalization showed its strongest additional accuracy benefits for inward reaching at short distances under moderate delay — a specific condition that suggests the benefit is real but bounded, not a universal cure.

For humanoid robotics teams building teleoperation pipelines — whether for [dexterous manipulation](https://humanoidintel.ai/glossary/dexterous-manipulation) data collection or remote deployment — this research offers a technically grounded framework for improving the quality of human demonstrations without waiting for zero-latency networks.

---

## The Problem: Uniform Gains Ignore How Humans Actually Move

Most production teleoperation systems apply a single motion scaling factor across all conditions. The operator moves the [end-effector](https://humanoidintel.ai/glossary/end-effector) input, the robot scales that motion by a fixed gain, and the system hopes for the best. This works tolerably at low latency. It degrades rapidly as delay increases, because visuomotor coordination breaks down asymmetrically — humans overshoot in some directions more than others, at some distances more than others, and the degree of overshoot is itself individual.

Jiang and Chua's approach reframes this as a fitting problem: given a participant's behavior profile across combinations of delay, direction, and distance, compute the scaling gains that minimize mean overshoot in simulation for each condition. The resulting parameter set is personalized — not just to the operator in the abstract, but to the specific spatiotemporal context of each subtask.

This is methodologically significant. Teleoperation research has historically evaluated mean performance across operators, which washes out individual variation. Running experiments with twenty participants and fitting per-participant models gives a cleaner signal on whether personalization actually adds value beyond what a good average gain would achieve.

---

## What the Data Actually Shows

Performance was assessed across five metrics: overshoot, endpoint error, trajectory smoothness, economy of motion, and a composite error-time metric. Motion scaling consistently improved performance relative to unassisted trials across all of these. The headline number — up to 20–25% gains in key metrics — is real, but the "up to" qualifier matters. Effects were condition-dependent, with longer delays producing the most pronounced improvements.

The personalization benefit specifically emerged for inward reaching at short distances under moderate delay. That's a narrow regime, and the paper is appropriately measured in its claims: personalization is described as demonstrating "additional accuracy benefits" in that condition, not as uniformly superior to a well-tuned uniform gain.

The evaluation on a telesurgical robot — beyond the virtual simulator — adds external validity. This isn't purely a simulation result. Whether the findings transfer cleanly to the embodiment configurations and control latencies typical of humanoid platforms is a separate question the paper doesn't address, but the telesurgical validation is a meaningful step toward real-world grounding.

---

## Why This Matters for Humanoid Teleoperation

The humanoid robotics industry is currently running two parallel bets on how to get robots to generalize: [imitation learning](https://humanoidintel.ai/glossary/imitation-learning) from human demonstrations, and autonomy stacks based on [Vision-Language-Action Models](https://humanoidintel.ai/glossary/vision-language-action-model). Both depend heavily on teleoperation data quality. If your demonstration data is contaminated by delay-induced overshoot and trajectory noise, your policy training starts from a degraded baseline — and [sim-to-real transfer](https://humanoidintel.ai/glossary/sim-to-real-transfer) becomes harder because the demonstrations themselves don't represent clean human intent.

This is the indirect but commercially important implication of Jiang and Chua's work. Better motion scaling doesn't just help the operator in the moment — it improves the quality of the trajectory data that feeds downstream learning systems. For teams collecting large-scale teleoperation datasets, even a consistent 10–15% reduction in endpoint error across operators could meaningfully shift data quality thresholds.

There's also a direct deployment angle. As humanoid platforms move toward remote operation in hazardous or inaccessible environments — a use case several well-capitalized teams are actively building toward — latency becomes unavoidable. Network delay over cellular or satellite links won't disappear. Personalized compensation frameworks like this one represent a software-layer mitigation that doesn't require hardware changes.

---

## Skeptical Notes

A few caveats deserve explicit acknowledgment.

First, the fitting procedure requires per-participant calibration. In a research setting with twenty subjects and a structured experimental protocol, this is tractable. In a production system where operators rotate shifts or where you need the compensation to work on day one with a new operator, the calibration overhead is a real deployment friction. The paper frames personalized scaling as "a foundation for more adaptive frameworks" — which is honest about the gap between current results and production readiness.

Second, the study used reaching tasks in a virtual simulator as the primary experimental environment. Reaching is a relatively constrained motion primitive compared to the multi-contact, whole-body manipulation sequences that define hard humanoid tasks. Whether direction- and distance-specific scaling generalizes to more complex trajectories — grasping, insertion, bimanual coordination — is unresolved.

Third, twenty participants is a meaningful sample for an HRI study but not large enough to make strong claims about population-level personalization effects. The variance in individual scaling parameters would be worth examining in a larger cohort.

---

## Key Takeaways

- **Personalized motion scaling delivers up to 20–25% performance gains** in delayed teleoperation compared to unassisted trials, based on a 20-participant study by Jiang and Chua (arXiv:2608.08200).
- **Uniform scaling gains are insufficient** — direction, distance, delay magnitude, and individual operator variation all require separate parameter fitting to maximize precision.
- **Performance improvements are most pronounced at longer delays**, the exact regime relevant to real-world network teleoperation.
- **Personalization showed specific additional benefits** for inward reaching at short distances under moderate delay — a bounded but real finding.
- **The indirect implication for humanoid data collection is significant**: cleaner teleoperation trajectories mean higher-quality imitation learning datasets.
- **Production deployment friction remains**: per-participant calibration requirements must be addressed before this approach can scale operationally.

---

## Frequently Asked Questions

**What is motion scaling in teleoperation?**
Motion scaling reduces the amplitude of an operator's hand movements before transmitting them to the remote robot, counteracting the tendency to overshoot targets when visual feedback is delayed. A scaling gain below 1.0 makes the robot move less than the operator's hand, improving precision at the cost of workspace range.

**Why does communication delay cause overshoot in telerobotics?**
The human motor system relies on continuous visual feedback to correct trajectory in real time. When that feedback is delayed, the brain's predictive models become miscalibrated — operators continue moving past the target before delayed visual confirmation of position arrives. The result is systematic overshoot that increases with delay magnitude.

**How does personalized scaling differ from standard motion scaling?**
Standard systems apply a single, fixed gain to all operators across all conditions. Personalized scaling fits separate gains per operator, per movement direction, per distance, and per delay level — capturing the individual and contextual variability that uniform gains ignore.

**Does this research apply to humanoid robots specifically?**
The study used a virtual simulator and a telesurgical robot for evaluation. The underlying principles apply to any teleoperated system subject to network delay, including humanoid platforms. However, direct validation on humanoid hardware with its specific kinematics and control latency profiles has not been demonstrated in this paper.

**What's the path from this research to production teleoperation systems?**
The primary gap is calibration scalability. Current results require a structured fitting session per operator. Future work the authors suggest would integrate adaptive frameworks that update scaling parameters online from contextual information — reducing or eliminating the need for explicit pre-deployment calibration.