Embedded Vision Hardware

Tactile-vision fusion supplants pure vision in robot perception

NeoteAI and Fudan TEAI's N0-VTLA model — the first vision-tactile-language-action model pretrained on tactile data at scale — achieved 63.8% mean success on a 20-task simulation suite versus 44.0% for the strongest vision-only baseline (π0.5), and swept all nine real-robot NeoReal tasks. The team built the full stack themselves: sensor hardware, data pipeline, model architecture, and benchmarks including NeoData and the self-developed visuo-tactile sensor. This is a decisive architectural signal: vision alone is insufficient for contact-rich manipulation, and companies that own proprietary touch-plus-vision sensor hardware will command the highest integration barriers.

Across the signal set, Intel RealSense cameras (D415, D435, D435i, D405, L515) appear as the default RGB-D input across no fewer than five independent robotics research pipelines — wrist-mounted manipulation, 6D pose estimation via FoundationPose, and physical AI benchmarking. This pervasive citation pattern is not coincidental: it reflects deep supply-chain lock-in at the component level, even as Intel's stock declined 33% since June 22 and the company broke a 45-year self-funding streak to raise external capital.

Why it matters · Intel's financial distress does not erode RealSense's embedded position in the near term, but creates an acquisition window for any strategic buyer seeking a dominant robotics perception platform.

The $300M seed round at a $1.1B valuation — backed by Nvidia, Samsung, Deviation Capital, Toyota, and Menlo Ventures — shatters conventional hardware fundraising norms and confirms that strategic investors are pre-empting rather than waiting for commercial proof. Separately, Even Realities raised $21M (seed, Innovation Endeavors and Xora) for AI-native smart glasses with ambient coding agents, and Elio's $21M Series A positions AI-native optical sensing as a critical supply-chain component.

Why it matters · The compression of hardware fundraising timelines toward software-style seed rounds means investors who wait for Series B proof points will face dramatically higher entry prices in embedded vision.

Samsung (5 deals) and Nvidia (4 deals) top the investor leaderboard for this theme — and crucially they are co-investing together, appearing jointly in the $300M seed round alongside Toyota and Deviation Capital. Samsung simultaneously leads global AI patent filings (682 U.S. filings in 2025, ahead of Google's 671) while also participating in a $21.7B growth-stage deal, signaling vertical integration ambitions from chip to device to perception stack.

Why it matters · When the two dominant compute platform companies co-anchor embedded vision rounds, it narrows the field of viable exit paths and raises the strategic premium on remaining independent players.

Elio's $21M Series A, framed explicitly as 'adaptive optical sensing for AI applications,' and the PostRound commentary calling AI-native programmable optical sensing 'a distinct AI hardware layer for machine perception that may become a critical supply-chain component as physical AI scales,' together mark the crystallization of a new product category distinct from both traditional image sensors and general compute. This is reinforced by Even Realities' Terminal Mode on smart glasses, which routes ambient AI agent outputs through an always-on optical interface.

Why it matters · Investors should track optical sensing as a separable line item in the embedded vision stack — category formation at the component level typically precedes a wave of M&A as platform vendors seek to own the sensing layer.