September 2026 brought two kinds of news in robot perception: hardware that shipped and benchmarks that exposed what perception still gets wrong. On the hardware side, GeoCue put Ouster's Rev8 OS1 Max colour lidar into drone survey payloads, RoboSense confirmed its tactile "skin" sensors enter commercial delivery this month, and Hesai passed 200,000 cumulative robot-lidar deliveries.
On the research side, three new arXiv releases — Bench2Dex, BRAVE-6D and EventEgoHands++ — all attacked the same weak spot: robots that see well enough to plan, but not well enough to touch. The short version for anyone buying or building this quarter: robot-specific perception hardware is now a volume business with falling prices, tactile sensing is finally leaving the lab, and the most interesting depth camera announced this year is not purchasable until 2027. Below is what each release actually says, what it means for your stack, and where the numbers stop short of the marketing.
Table of Contents
- Robots, not cars, are now driving lidar volume
- Tactile sensing leaves the demo stage
- Colour lidar collapses two sensors into one on survey drones
- What this month's benchmarks say perception still can't do
- Announced versus buyable — check the ship date
- Frequently Asked Questions
Robots, not cars, are now driving lidar volume
RoboSense's interim results, published 27 August 2026, put a number on a shift the industry has been predicting for years. The company reported 719,200 lidar units sold in H1 2026, up 169.6% year over year — and its robotics-and-other segment grew 510.4% to 282,600 units. Robotics is no longer a rounding error beside automotive; it is roughly 39% of the unit count. Hesai's figures point the same direction from a different angle.
At CES 2026 the company said its JT-series 360° hyper-hemispherical robotics lidar had passed 200,000 cumulative deliveries and that it would double annual capacity to over 4 million units in 2026, supplying mowing robots, autonomous mobile robots (AMRs) and humanoids. What that means practically: the cost floor for 3D perception on a mid-volume robot keeps dropping, because two suppliers are now building capacity for millions of robot-grade units rather than tens of thousands. If your last bill of materials review priced lidar out of a product, the assumption behind that decision is a year stale. It also means supplier concentration risk is real — a small number of vendors now set the price of the sensor most mobile robots depend on.
Tactile sensing leaves the demo stage
The most significant single line in RoboSense's August filing was not about lidar. The company stated that its visual-tactile and MEMS tactile sensors have completed development, with the first enterprise-scale program entering commercial delivery by the end of September 2026 and reaching start-of-production in Q4 2026. Two sensor types are involved and they solve different problems. Visual-tactile sensors put a camera behind a deformable surface and read contact as an image, giving rich shape and slip information at modest update rates.
MEMS tactile sensors are small force-sensing elements that report pressure directly, at higher rates and lower cost per point. Robot "skin" has been a conference demo for a decade. A named customer program with a start-of-production date is a different category of evidence: it implies yield, calibration and durability have been worked out well enough to warranty. For anyone building manipulation hardware, this is the month the question shifts from whether tactile sensing is available to whether your control stack can use it.
Colour lidar collapses two sensors into one on survey drones
GeoCue announced on 1 September 2026 that it had integrated Ouster's Rev8 OS1 Max native-colour digital lidar into its TrueView 1 drone payloads, showing the result at Commercial UAV Expo in Las Vegas from 1–3 September. The sensor is Build America, Buy America compliant — a procurement requirement that decides eligibility on US federally funded surveying contracts, regardless of technical merit. The specifications are what make the integration interesting. The Rev8 OS1 Max is rated at 200 m range at 10% reflectivity, with ±0.25 cm (1σ) precision and ±1.25 cm accuracy — roughly double the Rev7 generation — and it captures colour alongside the 3D points.
The 10% reflectivity figure matters: it is the hard case, roughly dark asphalt or wet vegetation, so it describes range you can count on rather than range against a retroreflector. Native colour removes a whole class of integration work. A conventional colourised point cloud needs a separate camera, a rigid mount, extrinsic calibration between the two, and time synchronisation — and every one of those is a failure mode in the field. One sensor producing coloured points removes the calibration step entirely and cuts payload weight, which on a drone converts directly into flight time.
What this month's benchmarks say perception still can't do
Three arXiv releases in September targeted the gap between seeing and doing. Bench2Dex released a benchmark of 26 bimanual manipulation tasks across 12 dexterous hands, with roughly 1,300 human-teleoperated demonstrations, and evaluated ACT, Diffusion Policy, pi0.5 and GR00T N1.5 against it. Read the limitation before you read the leaderboard: its tactile signal is a simulated contact-geometry image, not any real sensor's output — so a policy that scores well there has not been tested on the noise, drift and hysteresis a physical tactile sensor produces. BRAVE-6D, posted 15 September, benchmarks *active* vision in 6-DoF pose estimation — meaning the robot chooses where to look next, rather than scoring a fixed image. It uses Gaussian-splat view synthesis to render candidate viewpoints, which makes a next-best-view decision scorable. The target is the detect-and-grasp failure mode on small objects, where a single camera angle leaves the pose ambiguous and the grasp misses.
EventEgoHands++, also dated 15 September, released an event-camera method and a real dataset for egocentric 3D hand-mesh reconstruction. Event cameras report per-pixel brightness changes instead of frames, which handles motion blur and harsh lighting well but floods the stream with background events whenever the wearer's head moves — a problem this work addresses directly. That matters for teleoperation and for learning-from-human-demonstration pipelines, where the training signal is a human hand recorded from a head-mounted camera. The editorial framing came from *Frontiers in Robotics and AI*, which on 3 September published a collection of five papers on robotic perception, multi-robot exploration and mapping. It named two open problems: decentralised coordination under limited communication, and semantic scene understanding that generalises across domains. Both are honest descriptions of where fleet deployments break.
Announced versus buyable — check the ship date
The clearest trap in September's news is RealSense's D585 Pro. The SoC depth camera, shown at Automate 2026, delivers 1280×960 depth at 60 fps with sub-15 cm standoff and a claimed 2× depth-quality improvement over prior RealSense models — genuinely strong numbers for close-range bin picking and manipulation. It does not ship until Q1 2027.
Before you write a sensor into a design this quarter, separate the announcements into three buckets: Two further checks are worth making before committing. Confirm procurement compliance early if any US federal funding touches the project, because Build America, Buy America eligibility is a yes/no gate that no amount of performance closes. And when a spec sheet quotes range, confirm the reflectivity it was measured at — the Rev8's 200 m is stated at 10%, and a number quoted without that qualifier is not comparable to it.
- **Shipping now** — Ouster Rev8 OS1 Max in GeoCue TrueView payloads; Hesai JT-series at volume. Order against a quote.
- **Delivering within the quarter** — RoboSense tactile sensors, with commercial delivery stated for end of September and start-of-production in Q4 2026. Ask for a program reference, not a datasheet.
- **Announced only** — RealSense D585 Pro, Q1 2027. Fine to design around, dangerous to schedule around.
Frequently Asked Questions
Is tactile sensing actually purchasable yet, or still a research item?
RoboSense states its visual-tactile and MEMS tactile sensors have completed development, with a first enterprise-scale program entering commercial delivery by the end of September 2026 and start-of-production in Q4 2026. That is a named customer program rather than an open catalogue listing, so expect to negotiate rather than order.
Why does Build America, Buy America compliance come up in a lidar announcement?
US federally funded infrastructure and surveying contracts require BABA-compliant equipment. GeoCue's TrueView integration of the Ouster Rev8 OS1 Max is explicitly BABA compliant, which makes those payloads eligible for federally funded survey work that non-compliant sensors cannot bid on.
Can I trust a manipulation policy's Bench2Dex score on a real robot with real tactile sensors?
Not directly. Bench2Dex's tactile input is a simulated contact-geometry image rather than a physical sensor's output, so its scores measure policy behaviour under clean contact data. Real tactile hardware adds noise, drift and temperature effects the benchmark does not model.
What is an event camera, and why use one for hand tracking?
An event camera reports per-pixel brightness changes as they happen instead of full frames at a fixed rate, giving very high effective temporal resolution and good tolerance of motion blur and difficult lighting. EventEgoHands++ applies this to egocentric 3D hand-mesh reconstruction and handles the dense background events a moving wearer generates.
You Might Also Like
- What Is New With Tutorials and DIY Robotics in September 2026? Latest company releases and research papers and Key Takeaways
- What Is New With Robotics Companies in September 2026? Latest company releases and research papers and Key Takeaways
- What Is New With Mining Robotics in September 2026? Latest company releases and research papers and Key Takeaways
