Physical AI in 2026 has no shortage of grand narratives. Jensen Huang said the words “physical AI” seventeen times in one CES keynote; global funding topped $6.4 billion in a single quarter; every player from NVIDIA’s GR00T N1.6 to XPeng’s X-Foresight and Geely’s WAM is scrambling for position on the same crowded map. Meanwhile an awkward fact keeps fermenting: robots that backflip and dance on stage walk into a real kitchen and knock over the salad bowl.
The essence of the failure is not that the robot isn’t smart enough. It is that it cannot reliably understand what is in front of it.
From June 27 to 29, a Chinese AI company that rarely makes headlines — Hangzhou-based Om AI (Lianhui Tech) — released three models over three consecutive days, the VLX series. Three launch events, three models, three plain engineering goals: continuous perception, precise localization, action execution. And underneath the lineup sits a counterintuitive claim: physical AI’s most urgent bottleneck is neither the reasoning depth of the “brain” nor the predictive power of “imagination,” but the most basic stage of all — whether AI can observe the world continuously and actively, the way a person does.
The Perception Blind Spot Everyone Skips
The technical map of physical AI is no longer blank. VLA models handle language and decisions; world models simulate the consequences of actions and validate feasibility. Both mainlines are converging fast, and the industry is already debating their fusion. But all the attention sits on “deciding” and “predicting,” while everyone quietly skips a more upstream question: AI’s perception is, at its foundation, intermittent.
Mainstream multimodal models still process video like flipping a picture book: grab a frame, analyze it, answer once; when the next frame arrives, start over. Every sampled frame breaks the model’s context. If a user asks a question mid-video, the model answers from the current frame alone, knowing nothing about what happened before. This frame-flipping perception works fine for digital chat — the user asks, the model answers, the user always initiates. The physical world does not pause. The drone is flying, the robot dog is walking, the camera keeps recording. AI has no right to say “hold on, let me compute,” because between any two frames something important can happen — an obstacle appears, a tool gets moved, a person walks into a danger zone.
If perception itself is discontinuous, no amount of decision intelligence helps: you are correctly processing the wrong information, and the result is still wrong. That is the problem VLX set out to solve — and the easiest part of it to underestimate.
The VLX Trio: From Continuous Seeing to Direct Acting
VLX is not one model but a three-link capability chain — VLX-Flow, VLX-Seek, VLX-Go — mapping to the three baseline skills of any physical-world agent: see continuously, find precisely, act directly.
VLX-Flow keeps the AI genuinely online. The name is apt, because the core is not a stronger vision encoder but a change in how the model consumes information: from frame sampling to streaming. Video is chopped into continuous small chunks; each chunk gets incrementally encoded as it arrives and updates the model’s internal memory. The memory has two layers — a visual cache holding recent frame detail, and a semantic memory maintaining long-range context. When a question comes in, the model answers from maintained memory instead of replaying the whole video. Single-stream latency runs as low as 0.06 seconds, and multiple streams don’t pile up, because incremental update compute is constant regardless of how long the video runs. Traditional video understanding scales linearly with duration — more frames sampled, more recomputation, slower answers — tolerable for offline analysis, dangerous for real-time devices where every millisecond can be a wall you hit or a target you miss. The core innovation is not a benchmark-topping algorithm; it is a plumbing decision: the model stops waiting for a question to load context and instead stays online maintaining it. That decision separates a phone app from a physical device’s brain.
VLX-Seek selects regions instead of guessing coordinates. Even a continuously perceiving model hits a second problem: it knows something is nearby but cannot say exactly where. The traditional approach makes the model regress coordinates directly — X, Y, width, height — which is fundamentally “guessing numbers on a picture,” stable only in proportion to training data and notoriously brittle in small-sample regimes. VLX-Seek changes the task: instead of emitting coordinates, it picks the correct region from a set of predefined candidates — converting continuous regression into discrete retrieval. The difference is not an implementation footnote. It is the difference between a footballer taking a blind shot with no goal in view and the same player, facing the goal, choosing a corner to aim at. Candidate quality caps the ceiling, but choosing is far more stable than generating. The result is direct: higher localization accuracy from a smaller model with less data. On edge devices that is decisive — the embedded chips in industrial drones and inspection robots cannot host tens-of-billions-parameter giants, but a well-designed 0.6B-10B model paired with region retrieval delivers stable precision. The industry takeaway is worth sitting with: we may be too enchanted by “big and complete.” Sometimes “small and accurate” is worth more.
VLX-Go outputs waypoints, not advice. After perceiving and locating comes acting. Traditional VLA models emit text — “move left,” “forward two meters” — which a control system then translates into motor commands, and that translation layer leaks uncertainty at every step. VLX-Go skips it: the model directly outputs short-horizon waypoints, coordinate trajectories a robot can execute as-is, replacing “text suggestion plus control interpretation” with “waypoint prediction plus offline trajectory learning plus online RL.” Its parameter count is 0.6B, and deliberately so: it does not do deep reasoning, only real-time response — which means it knows what it can do and where its boundary sits, and stays in its lane.
All three models start from the same premise: edge hardware conditions are non-negotiable — limited compute, constrained power, latency that must be deterministic. Facing those constraints, the team refused to “compress a cloud model until it fits” and instead designed native architectures that maximize capability inside the edge boundary.
Edge-Native Versus Cloud-Compressed: Two Philosophies of Physical AI
This exposes the core fork in physical AI deployment today. Route one, cloud compression: train a giant on a big cluster, then distill, prune, and quantize it down to edge size. The ceiling is high — the cloud model reasons hardest and knows the most — but compression inevitably bleeds capability, and by the time the model fits an edge device it tends to “seem to understand everything while failing at everything on site.” Route two, edge-native: design the architecture inside the edge compute ceiling from day one — this much compute, this power budget, this latency allowance — and maximize capability within the boundary. The ceiling is capped by hardware, but stability is exceptional: no dependence on cloud compute or bandwidth means every on-site decision is local, immune to network drops and latency jitter.
Om AI chose route two, and it is no secret: CEO Zhao Tiancheng has described the company’s product as a “general physical AI brain” that different terminals access through a unified protocol — plug and play, like USB. The framework travels well beyond vision models: edge computing, IoT, autonomous-driving perception stacks, even mobile app development all face the same fork — build big and compress down, or optimize within the boundary from the start. Neither is absolutely right. When compute is abundant, bandwidth is cheap, and latency tolerance is generous, cloud compression wins on efficiency. But when the physical world’s three constraints stack — continuous time, dynamic environments, capped edge compute — the value of edge-native design has been badly underpriced.
An Inverted Priority in the Industry
Re-examine the industry’s technical bets through this framework and an odd pattern emerges. Nearly every headline player is building heavy cloud-side reasoning — NVIDIA’s Alpamayo, the compute substrate under Tesla’s Optimus line, World Labs’ spatial intelligence — projects that burn hundreds of GPU-hours answering “how smart can AI be in the physical world?” Om AI instead answered a more basic question: can AI hold its post stably in the physical world at all?
There may be an inverted priority here. Physical AI demos can lean on preset choreography, controlled environments, and abundant cloud compute. Real deployments — a factory drone, a subway construction-site patrol robot, a cross-region inspection system — treat network disconnection as normal, compute ceilings as hard, and latency uncertainty as fatal. At that layer, edge-native value is severely underestimated: not every terminal needs deep reasoning. An industrial inspection robot does not need to understand complex human behavior; it needs to reliably spot anomalies along a fixed route, mark locations, and trigger alarms. A 0.6B model that does that well may carry more industrial value than a 100B model’s lab demo.
If you build endpoint devices — drones, robots, security cameras — refuse to settle for cloud-model demos. Test edge inference latency stability directly: if your product cannot decide until footage round-trips to the cloud, it is locked to controlled environments forever. If you build vision models, consider the edge-native road: designing inside the hardware boundary first and maximizing capability second is often more practical than building massive and compressing later. If you invest in physical AI, look past parameter counts and stage demos, and ask one plain question: does it still work with no network? If the answer is no, it belongs in a launch keynote, not in the physical world.
Physical AI is still a marathon with the field unsettled. Some are betting on the depth of the brain’s reasoning; some on the predictive power of imagination. Om AI’s choice poses a humbler question: before you start thinking, can you first see the world in front of you clearly? For physical AI, perception that is always online may be worth more to industry than reasoning that is occasionally brilliant.
Sources: 36Kr coverage of Om AI’s VLX launch; the VLX-Flow, VLX-Seek, and VLX-Go repositories on GitHub (om-ai-lab); Xinhua reporting on Lianhui Tech.
