Humanoid Collision Avoidance Needs Body Sensors

That is why a recent Apple Machine Learning Research project called ARMOR is worth watching. The work is research, not a consumer product announcement, but the signal is practical: humanoid robots may need distributed, body-mounted depth sensing before they can move confidently through ordinary homes.

The ARMOR team reports that its wearable-like depth sensor layout reduced collisions by 63.7 percent and improved success rate by 78.7 percent compared with dense head-mounted and externally mounted depth camera setups. Against a cuRobo planning expert, Apple reports 31.6 percent fewer collisions, 16.9 percent higher success rate, and 26x lower computational latency. The real-world deployment used a Fourier GR-1, which makes the work especially relevant to the home humanoid category rather than only to warehouse AMRs or lab arms.

The buyer takeaway is simple: when a company shows a humanoid walking through a tidy demo room, do not only ask whether it has cameras. Ask where the cameras are, what parts of the robot body they protect, and whether the robot can plan quickly enough when the room changes.

Why Are Head Cameras Not Enough?

Most people understand robot perception as a face problem. If the robot has cameras in its head, it can see. If it has depth sensing or LiDAR, it can map the room. That is partly true, but it misses the awkward geometry of a humanoid.

A humanoid is tall, heavy, jointed, and constantly changing shape. Its shoulder, elbow, wrist, hip, knee, and foot all occupy different parts of space during a step or reach. A head-mounted camera can see the world ahead, but it may not see the side of a kitchen island near the hip, the edge of a dining chair near the knee, or the pet toy near the ankle at the right moment.

That matters more at home than in a lab. Homes are not motion-capture stages. They have glossy floors, moving people, loose rugs, open drawers, pets, children's toys, table legs, and narrow gaps between furniture. A robot that can walk across an empty studio may still be a poor fit for a small apartment if it cannot sense and protect the volume of its whole body.

This is also why "collision avoidance" is too vague as a buying claim. A small robot vacuum avoids obstacles by steering around them. A humanoid has to avoid collisions while balancing, swinging limbs, carrying payloads, and sometimes reaching into cabinets or over counters. The same phrase hides a much harder problem.

What Apple ARMOR Actually Tested

ARMOR stands for an egocentric perception system for humanoid collision avoidance and motion planning. Apple describes it as a hardware and software approach that uses wearable-like depth sensors on the robot, plus a transformer-based imitation learning policy trained in simulation using about 86 hours of human motion from the AMASS dataset.

The important part is not that Apple used the word "humanoid." It is that the comparison was about sensor placement and planning latency:

The caveat is just as important: this is a research result. It does not prove that any current home humanoid is ready to navigate a family room unsupervised. It does suggest a useful test for future product claims. If a humanoid company only talks about a head camera, a front depth sensor, or a neural network, the missing question is whether the robot has body-level spatial awareness.

The ui44 Database View

The home humanoid market is still split between research platforms, pre-order products, and development programs. The sensor story looks different across those groups.

There is no single winner in that table. 1X NEO is the most explicitly home-facing, with a $20,000 pre-order price, a softer 30kg body, tactile skin, and household chore positioning. That makes perception and safe contact central to its value. If NEO is going to work around furniture and people, body awareness is not a bonus feature. It is part of the core promise.

Unitree G1 is different. At $13,500, it is one of the most accessible humanoids in the database, but it is still best understood as a research and development platform. Its depth camera and 3D LiDAR help, and its lower 132cm height may reduce some household risk, but a buyer should not treat availability as proof of domestic collision safety.

Fourier GR-1 is the bridge between research and embodied hardware in this specific story because ARMOR was deployed on a real GR-1. In ui44's database, GR-1 is a 165cm, 55kg active humanoid platform with bipedal walking, stair climbing, payload carrying, force/torque sensors, IMU, cameras, and visual perception. That is exactly the kind of robot where whole-body collision avoidance matters: big enough to do useful work, but also big enough to damage furniture or hurt someone if motion planning is sloppy.

What Body-Mounted Sensing Could Change At Home

The best case for body-mounted depth sensing is not that it makes a humanoid smarter in the abstract. It is that it gives the planner fresher geometry around the parts of the robot most likely to collide.

In a kitchen, that could mean a robot notices a drawer handle near its hip while turning. In a hallway, it could mean the robot sees that its knee path is too close to a shoe rack. Near a dining table, it could mean the arm planner understands where the elbow will travel while lifting a plate.

It also changes what "safe around people" should mean. Human-aware navigation is not only about avoiding a person as a moving obstacle on a floor map. It is about a tall articulated machine knowing the volume of space its body will occupy during the next movement. A humanoid that can stop before stepping on a foot but still swings an elbow into someone is not truly home-safe.

The hard part is latency. More sensors can mean more data, more calibration, more failure modes, and more compute. ARMOR's 26x latency improvement versus a planning expert is interesting because home robots need quick reactions. A robot that spends too long planning may become frustrating, and a robot that skips planning may become unsafe.

What Should Buyers Ask Before Pre-Ordering?

For the next wave of home-facing humanoids, ask more specific questions than "does it have AI?"

Ask whether the robot has depth sensing only in the head or also near the torso, arms, hands, knees, or base. Ask whether the company has tested in cluttered rooms with furniture at multiple heights. Ask what happens if one sensor is blocked by clothing, a carried object, dust, or glare. Ask whether collision avoidance works while the robot is carrying something, not only while walking empty-handed.

For a product like NEO, the most important evidence would be home trials with repeated tasks: carrying laundry around furniture, opening cabinets near people, walking through narrow hallways, and reacting when someone unexpectedly steps close. For a platform like Unitree G1 or R1, the important evidence is different: developer documentation, sensor access, simulation support, and whether third-party teams can add or validate body-level perception themselves.

For expensive R&D machines such as Galaxea R1 Pro, the question is whether the platform's richer sensor stack and 360-degree LiDAR translate into manipulation safety in human spaces. It may have more perception hardware than a small humanoid, but it is also a 170cm dual-arm mobile manipulator with a much higher price and a research/industrial buyer profile.

What Could Go Wrong

Body-mounted sensors are not a magic fix. They add calibration requirements. They can be occluded by the robot's own limbs. They may struggle with reflective surfaces, glass, sunlight, or small moving objects. They also introduce cost and design trade-offs: every exposed sensor needs protection, wiring, compute, and a place on the body that does not interfere with motion.

There is also a privacy angle. More cameras and depth sensors around the body can capture more of the home. A useful home humanoid will need clear privacy controls, local processing where possible, visible recording states, and owner control over stored data. The safest robot is not only the one that avoids furniture. It is the one that does not turn a home into an uncontrolled sensor network.

Finally, sensor placement cannot rescue weak behavior design. A robot should slow down near people, avoid surprise movements, explain when it needs help, and fail gracefully. The best perception system still needs conservative product rules around children, pets, stairs, wet floors, clutter, and night operation.

Bottom Line

ARMOR is not a buying recommendation for any one robot. It is a signpost for how to judge the next generation of home humanoid claims.

If a humanoid is meant to share a real home, head cameras and demo videos are not enough. Look for evidence that the robot understands its whole body in space, plans fast enough for changing rooms, and has been tested around furniture and people at the body level. The most credible home humanoids will not merely see the room. They will know where their elbows, knees, hips, hands, and feet are about to go.