XR Daily: Pico Leaks Its Vision Pro Rival

Pico Space Pro Leaks Completely, Inside and Out

Japanese regulatory filings have now shown the Pico Space Pro from every angle. The documents sit under certification number 202-JWM032 at Japan’s Ministry of Internal Affairs and Communications, the local equivalent of the US FCC, and they contain both external photos of the assembled headset and full teardown shots of the interior: the front camera and sensor array, the optical modules, and a central cooling fan. UploadVR spotted the filings first, and Road to VR pulled the images out of both PDFs.

The model number is B3110, which is the same one that appeared in US FCC filings in September under the label “Mixed Reality Computer,” alongside a separate entry for the tethered unit described as a “Mixed Reality Main Unit.” That pairing tells you what Pico is actually building: a headset that does not carry its own compute and battery. It follows the same architectural bet as the Vision Pro’s external battery and every recent attempt to move heat and mass off the user’s face, an argument we broke down in our look at XR’s compute offload problem. A fan in the visor and a puck on the hip is a very different thermal budget than a standalone all-in-one.

The specs Pico already confirmed are the interesting part. Micro-OLED panels at 4000 pixels per inch. A dual-chip layout with a self-developed coprocessor dedicated to computer vision and image processing, paired with a new main processor that Pico says delivers twice the CPU and GPU performance of the Snapdragon XR2 Gen 2 that powers the Pico 4 Ultra and the Quest 3. On the software side, Pico OS 6 brings a visionOS-like interface and the RealityKit-inspired Pico Spatial Engine. Splitting CV work onto a second chip is the pattern to watch here: camera feeds, hand and eye tracking, and depth processing all run continuously, and that is exactly the workload a main SoC cannot absorb without thermal consequences.

The timeline remains strange. Pico announced a launch event in August, cancelled it days later, and pushed the device to Q4 2026. The US filing’s short-term confidentiality expires February 28, 2027, which is a soft deadline rather than a shipping date. The hardware appears finished. The launch plan does not.

→ UploadVR and Road to VR

---

Steam Frame: The Queue Slips to 2027, the Tools Open Up

Valve says the Steam Frame reservation queue now stretches into 2027. Customers can check their expected window directly on the product page while logged in, and order emails are going out either by December or sometime next year, with the 2027 group getting timing details in January. Each region runs its own queue, so a European reservation can move faster than a North American one. Valve explicitly notes that nobody on a waitlist leapfrogs a reservation holder, and waitlists are only touched once a region’s reservation queue empties.

The other half of the news is software. SteamOS 0.4 has shipped to all Frame users with power efficiency gains of up to 12% during both streaming and local play, a real fix for a headset that reviewers measured at roughly one to one and a half hours of mixed use. Streaming color clarity improved, and Valve reduced the frequency of “Error 17,” the incompatible protocol or connection failure that hits PC VR streaming, though that specific fix requires opting into SteamVR 2.18.2 or later on the beta branch. Wireless Adapter connectivity problems in some countries are resolved, and the manual Wi-Fi channel list now matches the region. Comfort aside, which we argued is the headset’s actual product decision in our Steam Frame comfort piece, sustained battery life under streaming load is the number that decides whether people use it for long sessions.

Buried at the bottom of the changelog: Valve released Steam Frame CAD files and electrical interface information for non-commercial use. That is an open invitation for hardware mods, replacement faceplates, straps, mounts, and the kind of accessory ecosystem the Quest got from third parties but Valve hardware historically did not. It costs Valve nothing and hands modders a documented starting point.

→ Road to VR and the SteamOS 0.4 change

---

Epic Tells Its Own Devs to Stop Hand-Writing Code

A Bloomberg report says Epic Games delayed an unannounced shooter built on Marvel and Star Wars characters, originally targeted for March 2027, and instructed the team to rebuild its production model around AI tools. One director’s instruction to developers was blunt: “No more hand-written code.” Developers are expected to guide and edit AI agents instead, with Anthropic’s Claude named specifically. Epic communications director Jake Jones framed it as developers remaining the creative drivers while using AI to accelerate production.

Tim Sweeney confirmed the direction publicly within hours. “We have a huge pivot underway here at Epic, moving developers around to retool for maximum AI acceleration,” he wrote, arguing the goal is more output from existing teams rather than the same output from fewer people. The context is hard to ignore: Epic laid off more than 1,000 employees in March, roughly 20% of staff, after Sweeney said declining Fortnite engagement had the company spending far more than it earned. The company also found over $500 million in savings from contractors, marketing, and unfilled roles, and said those cuts were not AI-driven.

The technical groundwork already exists. Unreal Engine 5.8 shipped experimental MCP integrations that let language models drive editor workflows directly, and Epic positioned AI-assisted development as a pillar of Unreal Engine 6 next to Verse. Whether that becomes a productivity multiplier or a maintainability problem is the open question, and it lands in the same territory as our piece on why XR’s next wave of games is coming from modders rather than studios: small teams with AI leverage can move faster than large teams reorganizing around it.

→ 80 Level

---

OmniBody Cuts and Peels Soft Objects in Real Time

Two KAIST researchers, Jewoo Shin and Jinah Park, presented OmniBody at SIGGRAPH 2026’s Animation and Simulation poster session, and the results are worth reading closely. The method throws out connected meshes entirely and replaces them with voxel-aligned, oriented ellipsoidal particles. Local Extended Position-Based Dynamics constraints handle contact, bending, anisotropic elasticity, volume preservation, and fracture, so global behavior emerges without matrix assembly or solving a linear system. That is the key architectural choice: it looks conceptually like finite element analysis, since each particle is an element with locally computed properties, but it skips the expensive global solve that keeps FEM out of interactive frame rates.

Destruction is where the representation pays off. In a mesh-based pipeline, cutting or fracturing an object forces vertex duplication and remeshing, which is exactly the cost that makes real-time slicing impractical. OmniBody deactivates the contact constraint between the affected particles instead, an operation that parallelizes trivially. Collisions get easier too, because the particles themselves serve as the collision boundary, so body-on-body, body-on-particle-material, and self-collision all follow the same path.

The numbers: a 64,000-particle object simulates in roughly 3.1 milliseconds on an RTX 3080 Ti, and a demanding cutting test stayed under 10 milliseconds. Shin notes that exposed interior surfaces remain an open challenge, which matters if you want to cut into an object and have the cross-section look correct. For XR the appeal is obvious, since hand tracking plus freely deformable, destructible objects is the interaction model sandbox games and training sims have been chasing for a decade without the frame budget to support it.

→ 80 Level

---

RimWorld Leaves Unity for an In-House Engine

Ludeon Studios announced update 1.7 for RimWorld, and with it the Waylith Engine, the studio’s own in-house renderer and simulation core. The game looks and plays identically, but internal tests show performance improved by 2 to 4x. A stress-tested late-game colony ran 2.2x faster than 1.6, startup, map generation, world generation, and ticks-per-second all improved, and a modpack with 1,500 mods dropped from 14 minutes of load time to 6. Update 1.7 is already on the unstable beta branch.

Migration is unusually painless. Savegames move in both directions, so a 1.6 save loads on 1.7 and a 1.7 save still works if you revert. The overwhelming majority of mods run unchanged, and the yellow “potentially unsupported” flags are only triggered by the version number changing. Modders gain access to current C# language features and a codebase where complex mods are less likely to tank performance, which is the practical payoff of controlling your own runtime instead of waiting on a third-party engine roadmap. The catch is hardware: PCs older than roughly ten years may not run it at all.

Ludeon also confirmed something new is in development on Waylith, without saying what. That is the more significant signal. An in-house engine only pays for itself across more than one title.

→ Ludeon Studios

---

ManimGX Renders 3D Animation Through WebGPU

Academa’s ManimGX keeps Manim CE’s Python API and replaces the renderer with Rust on top of wgpu. The pitch is agent-friendly 3D video: an LLM writes ordinary Manim-style Python, and the Rust GPU core turns it into an MP4. On an M4 Pro, 36.2 seconds of 1080p60 video across five scenes rendered in 3.08 seconds, which the project measures at 90.6x faster than ManimCE, 11.5x faster than ManimGL, and 18.7x faster than Blender’s Workbench renderer from launch to finished file.

Two details make this more than a benchmark flex. First, the browser path: npm install manimgx runs Python in a Pyodide Web Worker and renders through WebGPU, so a scene embeds in a page with no render server, and Linux wheels bundle Mesa’s lavapipe for CPU Vulkan when no GPU is available. Second, the lighting is adapted from Google’s Filament, giving physically based materials, image-based lighting, ambient occlusion, bloom, and AgX tone mapping rather than flat shaded output. The same wgpu abstraction that powers this is the layer under WebGPU in the browser, which is the stack we compared against WebGL in our WebGPU vs WebGL guide. If a Python-to-GPU path with no LaTeX toolchain and a headless browser target is this cheap now, the tooling floor for generating 3D visuals from scripts keeps dropping.

→ GitHub

---

Quick Bites

• Terry’s Terraforming gets a date: UploadVR’s hands-on narrative adventure has a release date, giving Quest owners something to play in the quiet stretch before the holiday hardware wave. → UploadVR
• Meta commits to 100+ live sporting events: Meta says Quest will carry more than a hundred live sporting events next year, extending its immersive college football deal into a real programming slate. → Road to VR
• U.N. warns on neurotechnology: The United Nations issued a warning about technology that reads or infers mental state, a reminder that the biometric inputs XR headsets already collect sit on a regulatory edge. → The New York Times
• Apogee rebuilds Mozilla’s Orbit locally: A fully local, private reimplementation of Mozilla’s Orbit surfaced on Hacker News, and it is the kind of on-device architecture XR headsets need for anything privacy-sensitive. → GitHub