Best practices for IMU-to-camera extrinsic calibration on a moving robot?

[karen.chen3]

@charlesbianchi From what I've seen: 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. Kind of makes me think about how different this all looked even three years ago.

[edward.nelson]

@karen.chen3 I can speak to this a bit. 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. 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.

[rivera14]

I don't think that's quite right, for what it's worth. 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. 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.

[joseph.robinson]

Slightly off-topic, but related: 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. 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.

[lperez]

@joseph.robinson I'd push back on this a bit. 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.

[klewis]

@lperez This is a great summary, thanks. 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. 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.

[chloe_jack]

This matches something I went through recently. 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. 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.

[sarahbernard]

Just to be precise about one thing: 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. 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.

[rivera14]

I'd take that specific number with a grain of salt, honestly. 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.

[forgesve15]

Genuine beginner question - 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. 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.