There's a number that keeps robotics founders up at night, and it isn't the one on their term sheets.
When Avea Robotics came out of Y Combinator earlier this year—the Spring 2026 batch, to be precise—the two-person San Francisco outfit made a claim that sounded almost quaint in its specificity: their Sentinel platform could achieve latency as low as 10 milliseconds under local network conditions. Ten. For the uninitiated, that's the delay between a human operator moving their hand in virtual reality and a robot, potentially thousands of miles away, mimicking that movement.
CEO Ary Indarapu and his co-founder Vikram Vadrevu, an aerospace engineer who spent years writing satellite flight software for NASA and the Department of Defense, aren't chasing that number for bragging rights. In the unglamorous arithmetic of Physical AI—the effort to make robots that can actually do useful things in the real world—latency determines everything. It's the difference between a robot that responds like an extension of your body and one that feels like you're piloting it through molasses.
More to the point: it's the bottleneck limiting how fast the industry can gather the massive troves of training data that foundation models need to learn. And right now, everyone is starving for data.
When Models Meet Reality
The money flooding into robotics tells part of the story. Startups in the space raised $40.7 billion in 2025, accounting for roughly 9% of all venture capital, according to CB Insights figures published in January 2026. Within that surge, investment in so-called world-model companies—the ones trying to teach robots to understand and predict physical environments—jumped from $1.4 billion in 2024 to $6.9 billion in 2025.
Yet as industry observers noted in January, these models remain hobbled by a fundamental scarcity. There simply isn't enough diverse, real-world training data. Scale AI laid out the problem in stark terms last September: the combined Open X-Embodiment and DROID datasets—two of the largest publicly available collections—totaled around 5,000 hours of robot interaction data as of late 2025. That's orders of magnitude smaller than the datasets that trained GPT or Claude.
Teleoperation sits at the uncomfortable center of this crisis. When a human straps on a VR headset and takes control of a robot—seeing through its cameras, feeling resistance through haptic feedback, commanding its movements in real time—the system captures exactly the kind of rich, multimodal sensor-action data that learning algorithms crave. Some in the industry have started calling this the "100,000-year data gap," a deliberately absurd shorthand for how far robotics lags behind the internet-scale datasets that made the current AI boom possible.
The dominant strategy for closing that gap, as analyses from late 2025 through early this year have repeatedly emphasized, is teleoperated data collection at massive scale. Which brings us back to those 10 milliseconds.
The Tyranny of the Ping
Human perception research has established fairly unforgiving thresholds. Above 200 to 250 milliseconds of end-to-end latency, direct teleoperation starts to feel noticeably sluggish; depending on the task's complexity, some operations remain tolerable up to 330 or 370 milliseconds. An EEG study published last November found perception thresholds clustering in the 100-to-200-millisecond window—past that, your brain starts noticing the disconnect.
These aren't abstract numbers. They create hard physical constraints for any company promising to control robots "anywhere in the world," as Avea's pitch puts it.
Recent research has begun to demonstrate what's achievable when teams obsess over every millisecond. A February paper on humanoid teleoperation reported end-to-end latency around 50 milliseconds, fast enough to enable reactive behaviors like catching thrown objects. An open-source stereoscopic VR framework published in March benchmarked 50 to 80 milliseconds under optimized local network conditions, rising to 80-100 milliseconds over more constrained wireless links. Crucially, those researchers published full instrumentation to prove it.
Commercial vendors, naturally, have made bolder claims. Adamo put out a blog post in March detailing their "sub-40-millisecond" stack, backing it up with a demo that controlled a robot from San Francisco to London. Avea markets that "as low as 10 ms" figure for local network conditions alongside features like six simultaneous full-HD video streams and haptic feedback across robot types ranging from 6-degree-of-freedom arms to full humanoids.
The catch, as with any "fastest on the market" claim, lies in the fine print. Local area network versus wide-area network. Motion-to-motion measurement versus input-to-photon. And perhaps most critically: what happens when network conditions degrade, as they inevitably do in the real world.
When Physics Gets in the Way
Wide-area networks impose their own stubborn reality. Starlink, frequently cited for remote operations, typically delivers latency of 20 to 60 milliseconds on land in 2026, with best-case scenarios around 20-25 milliseconds—though it can spike to 80-120 milliseconds. Public 4G and 5G networks commonly add 100 to 300-plus milliseconds end-to-end under real-world conditions—though a demonstration by NTT DOCOMO and Keio University earlier this year showed stable high-fidelity robot teleoperation over commercial 5G using network slicing.
In other words: the last mile remains a problem. Or the last thousand miles, depending on where your robot is.
A Suddenly Crowded Field
Avea isn't exactly racing alone. Extend Robotics markets what it calls an "immersive VR teleoperation platform" for dexterous control, with similar "anyone, anywhere" positioning. Humanola offers VR teleoperation software for humanoids like the Unitree G1—available, somewhat remarkably, through RobotShop for $5,500. Academic projects like XRoboToolkit, which won Best Paper at SII 2026, are pushing cross-platform OpenXR standards for dual-arm manipulation.
The landscape also includes infrastructure plays that signal where the puck is headed. Serve Robotics acquired both Phantom Auto and Voysys last September, integrating Voysys's directional foveated streaming SDK into its Level 4 sidewalk delivery autonomy stack. Voysys had specialized in ultra-low-latency streaming that cuts bandwidth while preserving what they characterized as "low and safe latency." Serve had planned to deploy 2,000 delivery robots by the end of 2025, each potentially requiring remote monitoring or intervention over 4G and 5G networks.
Teleo addresses the construction and heavy equipment vertical with supervised autonomy and remote operation, signing partnerships with Hitachi—demonstrated at CONEXPO-CON/AGG this year—and multiple dealer networks. In defense, L3Harris secured contracts in January for T7 EOD robots with jam-resistant teleoperation links for the U.S. Navy and Marine Corps. A reminder that latency matters quite differently when communications are actively contested.
The Long Road to Autonomy
Multiple data points suggest the years ahead—2026 through 2028, at least—will remain heavily reliant on teleoperation plus shared autonomy while vision-language-action models mature. Hyundai and Boston Dynamics unveiled an all-electric Atlas humanoid at CES this year with production deployments targeted for 2028. Industry commentary from TechRadar in January noted that many humanoid demos remain teleoperated, and teleoperation will be necessary for training and safe deployments in the near term.
Gartner predicted in April that by 2030, half of new warehouses in developed markets will be designed as "robot-centric, human-optional"—a trajectory that implies massive demand for remote operations centers and the low-latency video infrastructure to support them.
Consider the scale we're talking about. The International Federation of Robotics reported 542,000 industrial robot installations in 2024. China alone operates more than 2 million units. SoftBank's agreement last October to acquire ABB's Robotics division for approximately $5.4 billion—expected to close sometime later this year—signals consolidation around industrial robotics and Physical AI infrastructure at meaningful scale.
What We Don't Know
The challenge for any company claiming breakthrough speed is verification, and here's where things get murky. As multiple research papers from the past two years have emphasized, meaningful latency benchmarks require published end-to-end measurement methodology, clear specification of network conditions, and operator task success rates under jitter and packet loss.
Avea's "as low as 10 ms" figure lacks public context about whether that's a LAN measurement or includes wide-area links. The gap between lab demos and production deployments has a way of widening when robots leave controlled environments. It always does.
There's also the VR hardware question, which is less glamorous but no less important. IDC data shows XR shipments grew 44.4% year-over-year last year, with 33.5% growth forecast for 2026. But the same period saw Meta Quest shipments decline versus 2024 and weak Apple Vision Pro sell-through—IDC estimated roughly 45,000 units in the fourth quarter of last year. Meta raised Quest headset prices in April due to AI-driven RAM shortages, shifting total cost of ownership calculations for any company planning operator fleets.
Regulatory complexity adds another layer, though perhaps not the one founders focus on first. The EU's AI Act opened public consultation on high-risk classification guidance in May, with obligations phasing in through this year and next. The EU Machinery Regulation, fully applicable January 20, 2027, addresses software integrity and networked control for remote supervisory functions. CPRA updates that took effect January 1 have increased litigation around pre-consent tracking in California, making biometric and video data particularly sensitive for teleoperation platforms collecting training data from homes or workplaces.
Perhaps the most revealing perspective came from Sergey Levine, the UC Berkeley professor and Physical Intelligence co-founder, in a podcast this past May. He emphasized the centrality of real-world data while noting that teleoperation, though standard, must be complemented by other modalities.
It's a measured take that captures the industry's current reality better than any pitch deck: teleoperation is both critical and insufficient. The necessary foundation for scaling Physical AI, yes—but not the final answer.
The companies solving low-latency remote control aren't just building better joysticks. They're building the infrastructure that determines how quickly robots can learn to work without them. Whether Avea's 10 milliseconds holds up under scrutiny, or turns out to be another optimistic lab measurement, matters less than the direction of travel. The race is on, and latency is the finish line nobody's quite crossed yet.
