Robotics

Foundation models become the central robotics infrastructure layer

The robotics stack is consolidating around foundation model backbones rather than task-specific controllers. Alibaba's Qwen team deployed 23 researchers to build Qwen-RobotManip, a VLA built atop Qwen-VL that ranks 1st on RoboChallenge with a 20% relative improvement over π0.5 in out-of-distribution settings. Generalist AI's body-agnostic pre-training paradigm pushes task success rates from ~50% to 90% through staged fine-tuning, and Physical Intelligence continues to validate the World Model approach as the 'brain' layer. Genesis AI emerged from stealth with a $105M seed co-led by Eclipse and Khosla explicitly to build a foundational model for robots, signaling that pure-play robotics FM infrastructure is now fundable at frontier scale.

Agility Robotics was acquired by Churchill Capital Corp XI for $2.5B and simultaneously installed Microsoft veteran Peggy Johnson as CEO, marking the first major SPAC-style exit in humanoid robotics. This follows a $935M growth round at a $5.5B valuation and a $150M pre-Series B at $1B for other humanoid players, confirming that strategic buyers are moving from observation to acquisition. With Amazon (13 deals), Google (12 deals), and Nvidia (28 deals) dominating the investor table, the transition from VC-led to strategic-led capital is accelerating.

Why it matters · Pure-play robotics VCs face increasing pressure as hyperscalers and public-market vehicles compress exit timelines and set valuation floors.

The VIA system demonstrated 100% success on long-horizon Rainbow assembly tasks using frontier vision-language models without any robot-specific fine-tuning, directly challenging the prevailing paradigm that robots require dedicated training data. This parallels Qwen-RobotManip's unified alignment framework enabling large-scale multi-source training. If zero-shot or minimal-shot approaches generalize, the moat around proprietary robot training datasets narrows significantly.

Why it matters · Companies like Human Archive and Protege AI that monetize robot training data collection face a structural headwind if frontier VLMs commoditize the need for task-specific datasets.

Dexterity-first companies including RLWRLD (RLDX-1), Daimon Robotics, and Mimic Robotics AG are racing to embed force, tactile, and multimodal sensing into manipulation pipelines. Linkerbot and DexHand are scaling dexterous hand hardware specifically to serve this sensing layer. The thesis is that tactile ground-truth data is the one input that vision-only VLMs cannot synthesize, creating a defensible data moat for companies that own the physical sensing stack.

Why it matters · Early movers in proprietary tactile datasets will have a durable training-data advantage that is structurally difficult for hyperscaler foundation models to replicate.

Unitree is described as having done 'pioneering work' moving from hardware-first to large-brain AI, while AgiBot operates large-scale imitation learning pipelines at commercial scale, and Qwen-RobotManip's 23-researcher team applies LLM scaling recipes directly to physical AI. Yuanli Lingji/RoboticX launched in early 2025 with native VLM pretraining on robot data, and Agile Robots is a founding coalition partner in Nvidia's Cosmos 3 platform. Chinese firms now compete across the full robotics stack — silicon, embodiment, and foundation models — not merely on hardware cost.

Why it matters · Western robotics incumbents and investors must now model Chinese competitors as full-stack AI rivals, not just low-cost hardware manufacturers.