SHANGHAI — At the 2026 World Artificial Intelligence Conference (WAIC), almost every embodied AI robot showcasing its operational capabilities performed inside a guarded enclosure. Nearby safety personnel kept their eyes fixated on three specific areas: near the feet, around the hips, and along the arm movement trajectories.
These spots share a common physical blind spot: they all fall outside the field of view (FOV) of the head-mounted primary cameras, obscured by the robot's own body.
This blind spot is not a flaw in artificial intelligence models. Even the most advanced end-to-end vision-language-action (VLA) models cannot infer unobservable, occluded spaces—it is a matter of fundamental geometry. Yet, for robots to safely navigate human-centric environments, they must first master this critical distance of 5 centimeters to 1 meter—detecting floor drops before stepping forward or sensing nearby people when turning around.
The Divergence in Safety Architecture
Industry leaders like Tesla and Figure rely heavily on vision, multimodal models, and end-to-end learning at the core of their control systems, attempting to connect all sensors and actuators via a unified neural network. While this approach excels at high-level scene understanding and generalization, it struggles with two fundamental issues: physical body occlusions and the heavy compute overhead required for high-frequency, low-semantics proximity checks.
This challenge has shifted the industry toward a functional division of labor. While AI models handle perception and task planning, a dedicated ranging channel guards the safety baseline at close quarters, enabling lower-level control systems to slow down, detour, or stop instantly. This independent pipeline delivers shorter response latency and predictable failure modes—vital requirements for commercial mass production and customer acceptance testing.
Transitioning from AI Models to System Engineering
Robot safety is transitioning from a model capability issue into a full-scale system engineering challenge. According to market research firm IDC, global shipments of humanoid robots surpassed 18,000 units in 2025, with TrendForce projecting that figure to top 50,000 units in 2026. As robots enter factories and commercial facilities, external physical barriers and safety supervisors will no longer be practical.
Updated international standards, such as ISO 10218 (2025) for industrial robots and the developing ISO/CD 25785-1 for dynamic mobile robots, explicitly require strict risk assessment, safety integration, and speed and separation monitoring.
However, onboard computing power, battery capacity, and thermal space remain strictly limited. Offloading basic proximity detection entirely to heavy vision models drains power, increases latency, and inflates hardware costs.
To address this gap, LiDAR developer Benewake has launched the NOVA 20, a solid-state near-field LiDAR module designed specifically for the 5 cm to 1 m range. As Dr. Yuan Li, Founder and CEO of Benewake, highlighted in a public interview:
"As the first unit among perception, prediction, and planning-and-control, LiDAR allows us to see the bigger picture from smaller details during the perception phase."
Consuming under 1 watt, weighing less than 5 grams, and measuring just 25 mm × 22 mm × 13 mm, the NOVA 20 acts as a non-contact tactile layer distributed across wrists, torsos, knees, or ankles.
Solving the Five Proximity Challenges
The NOVA 20 addresses five primary engineering pain points in mass-production robotics:
- Complete Coverage: Featuring a 40° FOV across 20 channel detectors, it monitors a fan-shaped zone with centimeter-level precision to prevent falls and lateral collisions.
- Deterministic Control: By outputting raw distance data directly rather than heavy image streams, it reduces latency and mitigates multi-device cross-talk.
- Compute Offloading: Low power consumption and minimal payload requirements free up primary computing resources for semantic reasoning and manipulation.
- All-Weather Adaptability: Capable of operating under 100 Klux ambient light (equivalent to direct sunlight) with a blind zone under 5 cm, it enables seamless indoor-to-outdoor transitions.
- Bill of Materials (BOM) Integration: Utilizing a standard UART interface, it allows manufacturers to cost-effectively deploy sensors across critical risk points during mass production.
Redundancy Over Singular Perception
The debate between pure-vision setups and multi-sensor fusion is reaching a consensus. Future robotic safety architectures will rely on hardware redundancy across multiple sensor types and control tiers. Large models will govern high-level understanding and task execution, while dedicated physical sensors defend safety baselines at the lowest level.
Reflecting on the relationship between intelligence and perception, Dr. Li noted:
"The emergence and growth of strong intelligence will inevitably require strong perception, but 'strong' never means farther, larger, or more complex. For the robotics industry, true strong perception means reliably getting safety right at the exact distance where it is needed most—using hardware that is small, energy-efficient, and affordable enough."
Comments