What's the biggest perception failure mode you've personally debugged?

[lbianchi]

Wanted to get this in front of people who actually know the space. Unitree's Dex3-1 dexterous hand packs around 33 pressure/tactile sensors per hand across the fingers and palm, capable of sensing pressure roughly in the 10g-2500g range - a useful reference point for what 'production tactile sensing' looks like right now. Tactile skin arrays have improved a lot, but 'good enough to matter' really depends on the task - coarse contact detection across a large area is fairly mature, while fine, high-resolution force distribution sensing (like a human fingertip) is still the harder problem. Happy to be told I'm wrong on any of this.

[barbara.jones]

Just to be precise about one thing: Proprioception (the robot's sense of its own joint angles, velocities, and forces) tends to get less attention than flashy vision systems, even though a lot of balance and manipulation failures trace back to proprioceptive noise or miscalibration rather than a vision problem. This whole thread is a good reminder how young this field still is.

[diego.moore6]

I don't think that's quite right, for what it's worth. Sensor fusion mostly earns its keep by covering for each individual sensor's weaknesses - vision struggles with occlusion and lighting, IMUs drift, force/torque sensors are noisy at low loads - fusing them gives a more robust estimate than any one source alone, independent of raw compute.

[barbara.jones]

@diego.moore6 Ran into exactly this myself. Event cameras (which report per-pixel brightness changes rather than full frames) are still more of a research curiosity than a production sensor for humanoids, mainly because the software ecosystem and processing pipelines around them are far less mature than for standard frame-based cameras. Estimating joint torque from motor current draw is cheap and requires no extra sensor, but it's less accurate than a dedicated torque sensor because it doesn't capture friction losses through the gearbox - good enough for coarse control, not always for precise force-controlled tasks. Makes me wonder how this looks in another five years.

[emilyperez]

@barbara.jones This is a great summary, thanks. Force/torque sensors near the ankle give a direct read on ground reaction forces, which is valuable for balance control, but they add cost, a failure point, and routing complexity right at a joint that already takes the most mechanical abuse.

[emily.kumar]

@emilyperez This matches what I've seen too. Multi-camera calibration drifts over time from thermal expansion, vibration, and mechanical wear, which is why production systems typically run periodic recalibration routines rather than assuming a one-time factory calibration holds forever. Latency between a perceived event (like a slip) and a corrective control response matters enormously for balance - even 50-100ms of extra perception latency can be the difference between a smooth recovery and a fall, which is part of why a lot of balance-critical sensing is proprioceptive rather than vision-based.

[mia_lars]

@emily.kumar I'd take that specific number with a grain of salt, honestly. LiDAR gives reliable, lighting-independent range data but is heavier, pricier, and gives sparser point clouds up close than stereo or depth cameras, which is why a lot of humanoids lean on stereo/depth cameras for near-field manipulation and reserve LiDAR (if present at all) for longer-range navigation. IMU drift over time (bias instability) is usually the real culprit behind slowly diverging state estimates, not noise - it's typically handled with sensor fusion against other references (visual odometry, joint kinematics) rather than trying to eliminate drift at the source.

[chenperez]

@mia_lars That's the official framing, at least - reality tends to lag a bit. SLAM in a working warehouse is harder than in a controlled lab mainly because the map keeps changing - pallets move, people walk through, lighting shifts near dock doors - so a lot of production systems lean on semi-static maps refreshed periodically rather than pure continuous SLAM.

[young58]

@chenperez Just to be precise about one thing: A minimum viable sensing suite for safe bipedal walking generally includes joint encoders, an IMU for orientation/angular velocity, and either force/torque sensing or accurate current-based torque estimation at the ankles - everything else (vision, tactile, LiDAR) adds capability rather than being strictly required just to stay upright. Depth sensing range and reliability both degrade outdoors in direct sunlight for most structured-light and active stereo cameras, since the ambient IR washes out the projected pattern - it's a real limitation for humanoids intended for anything beyond indoor, controlled environments. Kind of makes me think about how different this all looked even three years ago.

[robertmiller]

Tangent, but worth mentioning: Multi-camera calibration drifts over time from thermal expansion, vibration, and mechanical wear, which is why production systems typically run periodic recalibration routines rather than assuming a one-time factory calibration holds forever. Unitree's Dex3-1 dexterous hand packs around 33 pressure/tactile sensors per hand across the fingers and palm, capable of sensing pressure roughly in the 10g-2500g range - a useful reference point for what 'production tactile sensing' looks like right now.