Jensen Huang spent most of NVIDIA's GTC 2025 keynote on it. Ouster rebranded itself as a "Physical AI perception platform" in its Q1 2026 earnings call. Every major LiDAR manufacturer's marketing deck now includes a slide about embodied intelligence.
Physical AI — AI systems that perceive, reason about, and act in the physical world — is the label the industry has settled on for what used to be called "embodied AI" or just "robotics." The term matters less than the shift it represents. A robot running a classical SLAM stack and a hand-coded state machine doesn't need the same sensor data as a robot running a Vision-Language-Action (VLA) model that makes decisions based on a 3D understanding of its environment.
LiDAR is useful for Physical AI. But which specs matter, and why is the answer different from what most spec sheets suggest?
Why Physical AI Changes the LiDAR Equation
Digital AI — chatbots, image generators, code assistants — processes data that's already structured. Text is tokenized. Images are pixel grids. The data is clean, static, and forgiving of latency.
Physical AI processes sensor data from a world that doesn't wait. A forklift backing up at 1.5 m/s covers 75 centimeters in the 500ms between LiDAR frames if your sensor only runs at 2 Hz. That's enough distance to miss a person stepping into its path, or to build a point cloud that's already stale by the time the VLA model processes it.
The mismatch between sensor output rate and AI processing speed creates a perception debt. If your LiDAR produces a dense, complete 3D view of the environment at 10 Hz but your onboard compute can only run the perception model at 5 Hz, you're paying for sensor performance you can't use. Conversely, if your sensor produces sparse, gappy data at 20 Hz but the model needs dense point clouds, no amount of frame rate makes up for missing information.
Physical AI demands a match between three things: what the sensor captures, what the compute can process, and what the task requires. Spec sheets rarely frame it this way. They list maximum numbers. Engineers need to think in terms of minimum viable perception.
Point Cloud Rate: Density vs. Speed
Point cloud rate (measured in points per second) is the single most discussed spec in LiDAR marketing. It's also the most commonly misunderstood.
A 300,000 points/second sensor sounds three times better than a 100,000 pts/s sensor. But perception quality depends on point density at the target distance, not raw throughput. A sensor scattering 300K points across a 100m range will produce a sparse cloud at 50m. A sensor concentrating 200K points in a 50m range (like the Tantu M360, with 200 kHz output) will produce a denser, more useful cloud for indoor and warehouse robotics where most objects of interest sit within 25 meters.
For Physical AI, the useful metric is points per square meter at your operating range. A robot navigating a warehouse aisle needs dense coverage of shelves, pallets, and forklifts at 5–15m range. A robot inspecting a factory floor needs detail at 1–5m. The spec that matters is the one that delivers enough points at your actual working distance to feed your perception model reliably.
Frame rate matters too, but in a different way than most people assume. For classical SLAM, 10 Hz is often sufficient — the algorithm accumulates points over multiple frames. For VLA-based systems that make single-frame decisions ("is it safe to proceed?"), you need the frame rate to match or exceed the decision loop frequency. If your VLA runs at 10 Hz, your LiDAR should output at 10 Hz or higher.
Field of View: Why Gaps Kill Perception
Physical AI systems need to understand the full 3D space around them, not just what's in front. A human worker rounding a corner, a forklift crossing an aisle, a box falling from a shelf — hazards come from every direction.
This is where FOV becomes a non-negotiable spec.
A 360° horizontal FOV is table stakes for any mobile robot. The vertical FOV is where real differentiation happens. Most 2D LiDARs scan a single horizontal plane, leaving everything above and below that plane invisible. Stepping up to a sensor with a 70° vertical FOV (the M360's -10° to +60° range) captures ceiling structures, overhanging obstacles, and the top of shelving units — information that a 2D sensor simply cannot provide.
For Physical AI perception models that reason about 3D scene structure (not just obstacle detection on a 2D plane), vertical coverage is the difference between a model that understands "there's a shelf to my left" and one that understands "there's a shelf to my left, with a box overhanging the edge at 1.2 meters height that I'll clip if I proceed."
Scanning pattern matters too. Rotating LiDARs with repetitive scanning patterns create systematic blind zones — areas that never get illuminated by the laser, frame after frame. Non-repetitive scanning (like the M360's random-access pattern) eliminates these fixed blind zones over time, which means the perception model sees a more complete picture with each passing second.
Blind Zone: The Spec That Determines Whether Your Robot Is Safe
Near-blind zone — the minimum distance at which a LiDAR can detect an object — is arguably the most safety-critical spec for Physical AI in human-shared environments.
Consider a warehouse robot approaching a person. At 1 m/s closing speed and a 500ms reaction time, the robot covers 50 cm before it can respond. If the LiDAR's blind zone is 50 cm, the person is invisible until the robot is already too close to stop. That's not a perception problem. That's a safety failure.
The M360's 5 cm blind zone means the robot can detect a person or obstacle while still 50 cm away — giving 500ms of reaction time at 1 m/s, which is enough for a VLA model to evaluate the situation and command an evasive maneuver.
For industrial settings where robots operate near workers (ISO 3691-4 safety zones, human-robot collaboration cells), blind zone spec directly maps to minimum safe approach distance. Smaller blind zones mean smaller safety buffers, which means tighter integration and higher throughput.
Power Consumption: The Edge Compute Constraint
Physical AI runs on edge compute. NVIDIA Jetson Orin NX delivers up to 100 TOPS of AI performance but draws 15–25W. An Orin Nano manages 40 TOPS at 10–15W. These are tight power budgets, and every watt allocated to sensor power is a watt not available for AI inference.
A LiDAR drawing 15W consumes a significant portion of the robot's total power budget — especially on smaller platforms like AMRs and delivery robots that may run on 100–200W total. That's 7.5–15% of available power going to one sensor.
At under 4.5W, the M360 draws less than a third of what many competing 3D LiDARs consume. On a 12–32V DC bus, that's a negligible load. The practical implication: you can run a 3D LiDAR without derating your compute platform or shrinking your battery, which matters when Physical AI models are themselves power-hungry.
Latency: Sensor-to-Decision Pipeline
Total perception latency in a Physical AI system is the sum of: sensor acquisition time + data transfer + preprocessing (filtering, downsampling) + model inference + actuation command.
LiDAR latency contribution is typically the smallest part of this pipeline — most sensors output data within a few milliseconds of the last point in a scan. The bottleneck is almost always the AI model. But sensor latency compounds with model latency in a way that matters.
If your sensor has a 100ms scan-to-output latency and your model takes 50ms per inference, you're at 150ms total. At 1.5 m/s robot speed, that's 22.5 cm of travel before the robot can react. For a VLA running at 10 Hz with 50ms inference, the 100ms sensor latency means you're always working with data that's at least one inference cycle old.
Low-latency sensor output (sub-50ms) matters when you're pushing decision frequencies above 10 Hz, or when your robot operates at speeds above 1 m/s. For slower industrial AMRs running at 0.5 m/s, it's less critical.
Specs That Don't Matter (As Much)
Some specs get disproportionate attention in marketing materials but have limited impact on Physical AI perception quality.
Maximum range. A 200m range spec looks great on paper. But for indoor robotics, warehouse navigation, and most commercial Physical AI applications, nothing relevant happens beyond 50m. Paying for range you don't use is paying for laser power that could have been allocated to higher point density at closer ranges.
Number of channels. More channels mean more resolution in the vertical axis. 128 channels at 90° vertical FOV gives 0.7° between channels. 32 channels at 70° FOV gives about 2.2°. For most Physical AI applications — obstacle detection, SLAM, scene understanding — this difference is marginal. It matters for long-range automotive perception where you need to distinguish a person from a signpost at 100m. It matters much less at 15m in a warehouse.
Matching Specs to Physical AI Scenarios
Different Physical AI applications place different weight on each spec. Here's how the priorities shift:
| Scenario | Point Rate | FOV | Blind Zone | Power | Range |
|---|---|---|---|---|---|
| Warehouse AMR | High | Critical | Critical | High | Low |
| Human-robot collab | Medium | Critical | Critical | High | Low |
| Outdoor inspection | Medium | High | Medium | Medium | High |
| Delivery robot | Medium | High | High | Critical | Low |
| Factory floor patrol | Medium | High | Medium | Medium | Medium |
For warehouse AMRs and human-robot collaboration — the two largest Physical AI deployment categories in 2026 — blind zone and FOV are the non-negotiable specs. Power consumption becomes critical when you're running VLA models on battery-powered platforms with limited thermal headroom.
The Sensor-Compute Balance
The practical advice for Physical AI teams selecting a LiDAR: stop optimizing individual specs in isolation.
Start with your compute budget. If you're running on a Jetson Orin Nano, your perception model can process roughly X points per second at Y Hz. Your LiDAR should produce point clouds that match or slightly exceed what the model can consume. Producing 500K pts/s that the model downsamples to 200K before inference is wasted sensor bandwidth.
Then match your FOV and blind zone to your operational environment. A 360° × 70° FOV with 5cm blind zone (like the M360) covers the full hemisphere that matters for indoor and near-outdoor robotics, with no near-field gaps that would create safety blind spots.
Finally, check power. If your LiDAR, compute, and motors together stay within your battery and thermal budget at sustained load, you're in good shape. If not, sensor power is usually the easiest thing to optimize — lower-power sensors exist that still deliver the point density and FOV you need.
Physical AI doesn't need the most expensive LiDAR on the market. It needs a LiDAR whose output characteristics match the requirements of the perception model, the constraints of the compute platform, and the safety demands of the operating environment.
Related Reading:
- M360 vs MID-360: Which 3D LiDAR Is Right for Your Robot?
- AGV LiDAR Comparison Matrix 2026: 12 Sensors Tested
- NDAA Section 164 LiDAR Ban: What It Means and How to Choose Compliant Sensors
Looking for a LiDAR sensor for your Physical AI robot? Contact our team to discuss your application requirements.