Every humanoid robot navigating a real warehouse or factory floor relies on a layered stack of sensors — cameras, LiDAR, depth sensors, tactile arrays — fused together in real time, but the specific mix each company chose reveals a genuine philosophical split: Agility's Digit and Boston Dynamics' Atlas both lean on LiDAR-based sensor fusion, Figure 03 pushes tactile sensitivity down to the weight of a single paperclip, and Tesla's Optimus skips LiDAR entirely in favor of a pure camera-based system that mirrors the same vision-only bet Tesla has made with its self-driving cars. How a robot senses its surroundings determines what it can safely do in a space full of people, shelves, and moving forklifts — and these approaches aren't interchangeable.
Agility Digit: Purpose-Built Warehouse Perception
Agility Robotics built Digit's sensor suite specifically around warehouse conditions rather than general-purpose navigation: LiDAR for precise 3D mapping, four Intel RealSense depth cameras for obstacle detection, an inertial measurement unit for balance, and dedicated force/torque sensors in each arm for grasp feedback. That hardware runs on a real-time Linux OS powered by dual Intel i7 processors. The software layer, Agility's Arc platform, is deliberately narrower in ambition than a general-purpose AI system: it excels at facility mapping, workflow definition, fleet management, and safety monitoring, and integrates directly with existing warehouse management systems — reliable and auditable, but not designed to autonomously learn entirely new tasks the way a broader AI model would. That focus has translated into real deployment: nearly 100 Digit units are working with clients including Amazon and GXO (serving Spanx among others), and the robot's practical operating rhythm — a roughly 4:1 work-to-charge ratio, meaning about four minutes of work per one minute of charging — reflects genuine engineering for continuous warehouse shift work rather than demo-length operation.
Figure 03: Tactile Sensitivity Down to a Paperclip
Figure 03 takes a different emphasis, leaning hard into vision breadth and tactile precision. It carries six stereo cameras offering a 60% wider field of view per camera than its predecessor, paired with fingertip tactile sensors sensitive enough to detect forces as small as 3 grams of pressure — Figure's own comparison point is that this is enough sensitivity to feel a single paperclip. Combined with 44 degrees of freedom and a 300-minute runtime enabled by inductive wireless charging, Figure 03's sensing approach is built around fine manipulation awareness as much as spatial navigation, consistent with its confirmed deployment on BMW's Spartanburg production line, where component-level precision matters as much as moving safely around the factory floor.
Tesla Optimus: Skipping LiDAR Entirely
Tesla's approach stands apart from every other robot covered here in one specific, deliberate way: Optimus has no LiDAR at all, relying instead on a pure vision system built around 8 cameras processed by a computer adapted from Tesla's Full Self-Driving stack. That data feeds end-to-end neural networks that translate visual input directly into motor commands, without an intermediate LiDAR-based 3D map the way Digit or Atlas build one. This is the same vision-only philosophy Tesla has publicly defended for years in its self-driving cars, where the company has argued cameras alone can match or exceed LiDAR-assisted perception given enough training data and compute — a bet Tesla is now extending directly into its robotics program. Tesla's Dojo supercomputer gives it a real training-data advantage at scale, purpose-built for processing video data faster than any other humanoid program can currently match, though as covered in an earlier entry in this series, that theoretical AI ambition still runs well ahead of confirmed, independently verified autonomous deployment.
Boston Dynamics Atlas: Full Sensor Fusion Plus Google DeepMind AI
Atlas takes the most sensor-rich approach among the widely covered humanoids, combining LiDAR, stereo cameras, and depth sensors for what Boston Dynamics describes as 360-degree environmental awareness, paired with a four-digit gripper carrying tactile sensing across both fingers and palms. Atlas's AI decision-making layer is co-developed with Google DeepMind — the same Gemini Robotics partnership covered in an earlier entry of this series — combining Boston Dynamics' hardware and sensor-fusion expertise with DeepMind's broader AI research. That combination positions Atlas at the premium, sensor-heavy end of the current field, with Boston Dynamics itself describing 2026 production as fully allocated, pushing new orders into 2027.
The Real Tradeoff: Narrow and Proven vs. Broad and Ambitious
Across all four approaches, the same fundamental tradeoff keeps showing up. Agility's Arc software is reliable, auditable, and already generating warehouse revenue specifically because it's narrow — it doesn't try to be a general-purpose AI brain, and that restraint is exactly why Amazon and GXO trust it with live operations today. Tesla's vision-only, LiDAR-free approach aims for a more general, broadly capable AI system, but that ambition currently outpaces verified real-world deployment evidence, consistent with the autonomy gap covered earlier in this series. Figure and Atlas sit somewhere between those two poles: rich, multi-modal sensing aimed at both navigation and fine manipulation, backed by real if still-limited production deployments at BMW and Hyundai respectively. None of these approaches has definitively "won" as of 2026 — they represent genuinely different bets on how much sensing complexity buys real-world reliability versus how much a smarter, more unified AI system can substitute for it.
Navigation and Sensing Stack at a Glance
| Robot | Core sensors | Notable capability | Software philosophy |
|---|---|---|---|
| Agility Digit | LiDAR, 4x depth cameras, IMU, force/torque sensors | 4:1 work-to-charge ratio for continuous shifts | Arc: narrow, reliable, WMS-integrated |
| Figure 03 | 6 stereo cameras, fingertip tactile sensors | Detects forces as small as 3 grams (a paperclip) | General-purpose manipulation and navigation |
| Tesla Optimus | 8 cameras only — no LiDAR | End-to-end neural network, vision-only (like Tesla FSD) | Broad AI ambition via Dojo-trained models |
| Boston Dynamics Atlas | LiDAR, stereo cameras, depth sensors, tactile gripper | 360-degree environmental awareness | Co-developed with Google DeepMind (Gemini Robotics) |
Frequently Asked Questions
Does Tesla Optimus use LiDAR to navigate?
No. Unlike Agility's Digit, Figure 03, or Boston Dynamics' Atlas, Tesla Optimus relies entirely on an 8-camera vision system, mirroring the same vision-only philosophy Tesla has used in its self-driving cars, with no LiDAR sensor at all.
How sensitive is Figure 03's tactile sensing?
Figure 03's fingertip tactile sensors can detect forces as small as 3 grams of pressure, which Figure describes as enough sensitivity to feel a single paperclip, supporting fine manipulation tasks on BMW's production line.
Which humanoid robot has the most warehouse deployment experience?
Agility Robotics' Digit has the most established warehouse presence, with nearly 100 units deployed at clients including Amazon and GXO, running on a software platform specifically built for facility mapping, workflow management, and safety monitoring.
Is a robot with more sensors always better at navigation?
Not necessarily. Tesla's camera-only Optimus takes a deliberately different bet than the LiDAR-equipped Digit, Figure 03, and Atlas, aiming for broader AI capability rather than sensor redundancy, though that ambition currently runs ahead of verified independent deployment evidence.