Robot Sensor Landscape & OpenRAL Catalog Roadmap

This document is the canonical reference for which sensors are common on production and research robot platforms in 2024–2026, and which of them should ship as first-party SensorBundle factories under python/sensors/.

It is the input for sequencing work in python/sensors/ and for any future ADR that proposes extending the SensorModality enum in openral_core.schemas.

Method: a survey of vendor data sheets, model cards from major humanoid / quadruped / manipulator vendors (Tesla, Figure, Unitree, Boston Dynamics, Agility, ANYbotics, Franka, UR, Kuka, LeRobot SO-100/SO-101, Aloha), and the LeRobot, ROS 2, and PX4 driver ecosystems. Numeric fields below come from manufacturer data sheets and are nominal — for any deployment they must be replaced with calibrated values via openral calibrate.

See also: - python/core/src/openral_core/schemas.py — normative SensorSpec, SensorBundle, SensorModality, IntrinsicsPinhole. - python/sensors/src/openral_sensors/realsense.py — reference implementation pattern (RealSense D435 / D455 today). - CLAUDE.md §6.1 (layer discipline), §7.5 (working with sensors).

1. Snapshot — sensor stacks of representative 2024–2026 robots

Robot	Class	RGB	Depth/RGBD	LiDAR	IMU	F/T	Tactile	Other
LeRobot SO-100 / SO-101	manipulator	1× wrist + 1× scene USB UVC (Logitech C920 / Arducam)	—	—	—	—	—	joint encoders (Feetech STS3215)
Aloha (bimanual)	manipulator	4× USB (2 wrist + 2 scene)	—	—	—	—	—	leader-follower position
Franka FR3 / Panda	manipulator	wrist RGB(D) (often RealSense D405/D435i)	optional D435i / D405	—	optional	wrist 6-axis (Bota SensONE / Robotiq FT-300) or joint-torque-derived	optional GelSight / DIGIT	7× joint torque
UR5e / UR10e	manipulator	optional wrist cam	optional	—	optional	wrist 6-axis (Robotiq FT-300, OnRobot HEX-E)	optional	—
Stretch 3 (Hello Robot)	mobile manipulator	head pan-tilt RGBD (D435if)	D435if	RPLIDAR A1	onboard	—	—	wheel odometry
Boston Dynamics Spot	quadruped	5× mono fisheye	5× depth	optional Velodyne / Ouster on payload	onboard	leg position/force	—	Spot CAM+ adds PTZ + thermal
Unitree Go2	quadruped	HD wide-angle	optional Intel D435i	Unitree L1 (360°×90°) or Livox Mid-360 (Edu+)	onboard	foot-end force	—	—
Unitree G1	humanoid	head depth (Mid-360 variant) + body cams	yes	optional 3D LiDAR	onboard	optional	optional fingertip tactile	joint torque sensors
Tesla Optimus Gen 2/3	humanoid	8× monocular RGB	—	—	onboard	—	fingertip tactile (gen 3)	—
Figure 02 / 03	humanoid	6× RGB (incl. palm cams on 03)	—	—	onboard	—	fingertip force on 03 (≈3 g sensitivity)	—
Agility Digit	humanoid	torso depth + perception array	yes	top-mounted lidar (variant-dependent)	onboard	foot/wrist F/T	—	—
ANYbotics ANYmal	quadruped	RGB + perception	depth	Velodyne / Robosense (variant)	onboard	foot F/T	—	—
Skydio 2+/X10 (drone)	drone	6× RGB navigation	stereo-derived	—	onboard	—	—	GPS, barometer

This is a snapshot for prioritization, not a normative table — entries vary by hardware revision and customer-specific payload.

2. Catalog by modality

Each row maps cleanly to a SensorSpec unless flagged in the Schema notes column. "Bundle" means the sensor naturally lives in a multi-sensor SensorBundle (RGB + depth + IMU on a RealSense, etc.).

2.1 RGB cameras (`SensorModality.RGB`)

USB / UVC RGB cameras dominate the LeRobot ecosystem and most low-cost research arms. Industrial vendors (FLIR, Basler) appear on factory-grade platforms.

Vendor / model	Form factor	Native res @ fps	hFOV	Encoding	Used on	Schema notes
Logitech C920 / C920e / C922	USB UVC	1920×1080@30, 1280×720@60	~78° (C920)	`mjpeg`/`yuyv`/`rgb8`	LeRobot SO-100/SO-101 default, Aloha	clean fit
Logitech Brio 4K	USB UVC	3840×2160@30	65°/78°/90° (selectable)	`mjpeg`/`rgb8`	research scene cam	clean fit
Arducam IMX179 / IMX291 USB	USB UVC	1920×1080@30	varies by lens	`mjpeg`/`yuyv`	community SO-100/SO-101 wrist	clean fit
Raspberry Pi Camera Module 3	CSI-2	4608×2592@15, 1080p@50	66°/120°	`rgb8`	embedded research	clean fit
FLIR Blackfly S (Spinnaker)	GigE / USB3	model-dependent	model-dependent	`bayer_rggb8`/`rgb8`	factory cells	clean fit
Basler ace 2	GigE	model-dependent	model-dependent	`bayer_rggb8`/`rgb8`	factory cells	clean fit
GoPro Hero (HDMI capture)	action cam	4K@60 / 1080p@240	~120°	`rgb8`	wrist/head research	clean fit
Leopard Imaging IMX477	USB / MIPI	4056×3040@30	lens-dependent	`rgb8`	research stacks	clean fit

2.2 Depth & RGBD (`SensorModality.DEPTH`, plus RGB+depth `SensorBundle`)

Active IR stereo (RealSense, Orbbec Gemini) and ToF (Femto, Azure Kinect) dominate. Stereolabs ZED uses passive stereo (GPU required). Luxonis OAK-D adds on-board AI. Status reflects vendor activity in 2025–2026.

Vendor / model	Tech	Range (m)	Native depth res @ fps	hFOV	RGB included	Used on	Schema notes
Intel RealSense D435	active IR stereo	0.1 – 10	640×480@30 (up to 1280×720@90)	87°	yes	Stretch, research arms — already in `realsense.py`	bundle
Intel RealSense D435i	D435 + BMI055 IMU	0.1 – 10	640×480@30	87°	yes	Unitree Go2 Edu, Stretch — add factory	bundle (RGB+depth+IMU)
Intel RealSense D455	wider-baseline IR stereo	0.6 – 6	1280×720@30	87°	yes	research humanoids — already in `realsense.py`	bundle
Intel RealSense D405	short-range IR stereo	0.07 – 0.5	1280×720@90	87°×58°	yes (mono)	Franka wrist, dexterous manipulation — high priority	bundle
Intel RealSense D415	rolling-shutter IR stereo	0.16 – 10	1280×720@30	65°×40°	yes	legacy research	bundle
Intel RealSense L515	solid-state lidar (EOL)	0.25 – 9	1024×768@30	70°	yes	discontinued; legacy	bundle
Orbbec Gemini 335	active IR stereo	0.17 – 10	1280×800@30	87°×58°	yes	RealSense replacement — high priority (active vendor support)	bundle
Orbbec Gemini 336 / 336L	active IR stereo (long-range)	up to 20	1280×800@30	varies	yes	mobile robots	bundle
Orbbec Femto Bolt	ToF (Sony chip; ex-Azure-Kinect)	0.5 – 5.46 NFOV	1024×1024 NFOV	75°×65° NFOV / 120°×120° WFOV	yes (4K)	research humanoids; Azure Kinect successor	bundle
Orbbec Femto Mega	ToF	0.5 – 5.46	1024×1024	similar	yes	networked research	bundle
Microsoft Azure Kinect DK	ToF (EOL)	0.5 – 3.86 NFOV	1024×1024	75°×65° NFOV	yes (4K)	discontinued; legacy datasets	bundle
Stereolabs ZED 2 / 2i	passive stereo (GPU req.)	0.3 – 20	2208×1242@15	110°×70°	yes (stereo)	outdoor research, drones	bundle; needs `STEREO` mod + `baseline_m` (see §4)
Stereolabs ZED Mini	passive stereo	0.15 – 12	1280×720@60	90°×60°	yes	XR / wrist research	bundle
Stereolabs ZED X / X One	passive stereo (GMSL)	varies	2560×1440@30	varies	yes	NVIDIA Jetson + Orin platforms	bundle
Luxonis OAK-D / OAK-D Pro	active IR stereo + on-board AI	0.2 – 4 typical	1280×720@120	71°×56°	yes	research, drones	bundle; AI accelerator metadata
Carnegie Robotics MultiSense S21/S30	passive stereo + spinning lidar	varies	2048×1088@15	varies	yes	Spot CAM+, Atlas legacy	bundle

2.3 3D LiDAR (`SensorModality.POINT_CLOUD` with `n_channels`)

3D LiDARs are reported under POINT_CLOUD with n_channels set to the vertical-channel / scan-line count (or omitted for non-repetitive solid-state designs). The schema's range_min_m / range_max_m and fov_h_deg / fov_v_deg cover the rest. No new modality is needed.

Vendor / model	Tech	Range (m)	hFOV × vFOV	Channels / pts·s⁻¹	Used on	Schema notes
Livox Mid-360	hybrid solid-state, non-repetitive	0.1 – 40 (@10% reflectivity)	360° × 59°	n.r. / 200k	Unitree Go2 Edu+, AGVs, lawn mowers — highest priority	`n_channels = None`; note non-repetitive in metadata
Livox Mid-70	hybrid solid-state	0.05 – 90	70.4° × 4.5°	n.r. / 100k	research	metadata flag
Livox Avia	hybrid solid-state	up to 450	70° × 77°	n.r. / 240k	UAV mapping	metadata flag
Livox HAP	solid-state	up to 200	120° × 25°	n.r.	automotive	metadata flag
Ouster OS0-32 / OS0-64 / OS0-128	rotating digital lidar	0.3 – 50	360° × 90°	32/64/128 / up to 5.2M	Spot payload, ANYmal, B1	clean fit
Ouster OS1-32 / OS1-64 / OS1-128	rotating digital lidar	0.3 – 120	360° × 45°	32/64/128	mid-range mobile	clean fit
Ouster OS2-128	long-range	1 – 240	360° × 22.5°	128	autonomous platforms	clean fit
Velodyne VLP-16 (Puck)	rotating	1 – 100	360° × 30°	16 / 300k	legacy autonomous, research	clean fit
Velodyne VLP-32C / Alpha Prime	rotating	up to 200 / 245	360° × 40° / 40°	32 / 128	autonomous	clean fit
Hesai PandarXT-32 / XT16	rotating	up to 120	360° × 31°	32 / 16	mobile	clean fit
Hesai JT16 (2025)	hemispherical + circular	up to 30	360° × 75°	16	small robotics	clean fit
Robosense Helios 16/32	rotating	up to 150	360° × 31°	16/32	mobile	clean fit
Robosense Bpearl	hemispherical	up to 30	360° × 90°	32	quadruped, AMRs	clean fit
Robosense E1R (2025)	solid-state Flash	up to 30	120° × 90°	n/a	outdoor robotics	metadata flag
Unitree L1	rotating-mirror solid-state	0.05 – 40	360° × 90°	n/a	bundled with Go2 / G1	metadata flag

2.4 2D LiDAR (`SensorModality.LIDAR_2D`)

Vendor / model	Range (m)	Angular range	Rate (Hz)	Used on	Schema notes
SLAMTEC RPLIDAR A1	12	360°	5.5	Stretch, hobby AMRs — high priority (low-cost)	clean fit
SLAMTEC RPLIDAR A2M12	16	360°	10	research AMRs — high priority	clean fit
SLAMTEC RPLIDAR S1 / S2 / S3	40 / 30 / 40	360°	10 / 10 / 15	indoor AMRs	clean fit
Hokuyo URG-04LX-UG01	4	240°	10	research, low-power	clean fit
Hokuyo UTM-30LX	30	270°	40	outdoor mobile, legacy Atlas head	clean fit
Hokuyo UST-10LX / UST-20LX	10 / 20	270°	40	industrial AMR — high priority (industrial)	clean fit
SICK TIM-561 / TIM-781	8 / 25	270°	15 / 15	factory AMR	clean fit
SICK LMS1xx / LMS5xx	20 – 80	270°	25 – 50	factory AGV	clean fit

2.5 IMU (`SensorModality.IMU`)

accel_noise_density and gyro_noise_density cover the noise model. A 9-axis unit (with magnetometer) carries the magnetometer term in metadata until a separate modality is added (see §4 gap M).

Vendor / model	DoF	Rate	Application	Schema notes
Bosch BMI270	6	up to 6.4 kHz	embedded boards (PX4, RPi HATs) — medium priority	clean fit
Bosch BMI088	6	up to 1.6 kHz	PX4 drones, robotics	clean fit
InvenSense ICM-40609	6	varies	embedded inside Livox Mid-360	metadata only
InvenSense ICM-20689 / -20602	6	up to 8 kHz	Pixhawk-class flight controllers	clean fit
Xsens MTi-3 / MTi-30 / MTi-100	9	100 – 400 Hz	research mobile / outdoor	clean fit (mag in metadata)
Xsens MTi-630 / MTi-680G	9 + GNSS	400 Hz	outdoor humanoids, AMRs	bundle (IMU + GPS)
VectorNav VN-100 / VN-110	9	800 Hz	drones, gimbals	clean fit
VectorNav VN-200 / VN-300	9 + GNSS	400 Hz	UAV navigation	bundle (IMU + GPS)
Microstrain 3DM-CV5-25 / GX5-25	9	up to 1 kHz	mobile robots	clean fit
Microstrain 3DM-GQ7	9 + dual-RTK GNSS	1 kHz	outdoor autonomy	bundle
Analog Devices ADIS16470 / 16475	6/9	up to 2 kHz	industrial / mil-spec	clean fit
STIM277 / STIM300	9	1 kHz	high-accuracy industrial	clean fit

2.6 6-axis force/torque (`SensorModality.FORCE_TORQUE`, `n_axes = 6`)

Vendor / model	Mounting	BW (Hz)	Range	Used on	Schema notes
ATI Mini40 / Mini45	wrist (research)	7000	low	Franka, UR research	clean fit
ATI Gamma / Delta	wrist	7000	mid/high	industrial cells	clean fit
ATI Nano17	fingertip	7000	very low	dexterous research	clean fit
Bota Systems SensONE	wrist (UR/Franka kits)	800	mid	UR / Franka — medium priority	clean fit
Bota Systems Rokubi	wrist	1000	mid	research	clean fit
Robotiq FT 300-S	wrist (UR-native)	100	mid	UR — medium priority (UR ecosystem)	clean fit
OnRobot HEX-E QC / HEX-H QC	gripper-arm interface	500	mid/high	UR, OnRobot ecosystem	clean fit
Wittenstein WT 24 / 26	wrist	1000	mid	research	clean fit
(intrinsic) Franka Panda / FR3 joint torques	per-joint	1000	per-joint	Franka — covered by `JointSpec.has_torque_sensor`	covered
(intrinsic) Kuka iiwa joint torques	per-joint	1000	per-joint	Kuka — covered by `JointSpec.has_torque_sensor`	covered

2.7 Tactile (`SensorModality.TACTILE_VISION`, `SensorModality.TACTILE_ARRAY`)

Vision-based fingertips fit TACTILE_VISION; magnetic / capacitive / resistive arrays fit TACTILE_ARRAY with tactile_grid = (rows, cols).

Vendor / model	Tech	Resolution	Mount	Used on	Schema notes
GelSight Mini	vision-based (camera + gel)	~320×240 image, ~30 µm spatial	fingertip	Franka, UR research, dexterous hands — high priority	`TACTILE_VISION`
GelSight DIGIT	vision-based (open-source design)	640×480@60	fingertip	open-source dexterous research — high priority	`TACTILE_VISION`
GelSlim 4 / GelSight Wedge	vision-based (flat finger)	varies	parallel-gripper finger	research	`TACTILE_VISION`
AnySkin (NYU/Pinto)	magnetic, replaceable	5-axis force, ~32 magnets	hand skin	open-source dexterous — high priority	`TACTILE_ARRAY`; magnetic in metadata
ReSkin (CMU/Meta)	magnetic, replaceable	similar	hand skin	open-source	`TACTILE_ARRAY`; magnetic in metadata
XELA Robotics uSkin	magnetic	4×4 taxels × 3-axis force	fingertip / palm	dexterous research	`TACTILE_ARRAY` 4×4
Contactile PapillArray	barometric / capacitive	9 taxels × 3-axis	fingertip	research	`TACTILE_ARRAY` 3×3
PaXini DexH13	capacitive	varies	fingertip	humanoid hands	`TACTILE_ARRAY`
TacTip family	vision-based	varies	fingertip	research	`TACTILE_VISION`

2.8 Audio, GPS, battery — already covered

Modality	Common hardware	Schema notes
`AUDIO`	ReSpeaker 4 / 6-Mic Array, USB lavaliers, robot built-ins	clean fit
`GPS`	u-blox ZED-F9P (RTK), u-blox NEO-M9N, Septentrio AsteRx-i / mosaic-X5	clean fit
`BATTERY`	smart-battery telemetry over CAN/I²C; surfaced via `WorldState.battery_pct`	clean fit
`JOINT_STATE`	actuator encoders (Feetech STS3215 on SO-100/101, Dynamixel, Maxon, Harmonic Drive)	typically expressed via `JointSpec.has_position_sensor / has_velocity_sensor / has_torque_sensor`, not a standalone `SensorSpec`

3. Schema-compliance audit

For each modality above, the existing SensorSpec fields are sufficient (green) unless flagged. Summary:

Modality	`SensorModality` enum	Sufficient `SensorSpec` fields?	Notes
RGB	`RGB`	✅ `intrinsics`, `encoding`, `fov_*`	—
Depth	`DEPTH`	✅ `intrinsics`, `range_`, `fov_`	—
Stereo (passive)	`STEREO`	⚠️ baseline missing	put `baseline_m` in `metadata` until §4 gap S resolved
RGBD bundle	RGB + DEPTH (+ IMU)	✅ via `SensorBundle`	sync = `hardware` for RealSense / Orbbec
2D LiDAR	`LIDAR_2D`	✅ `range_*`, `fov_h_deg`, `rate_hz`	—
3D LiDAR	`POINT_CLOUD` + `n_channels`	✅ for rotating; metadata for non-repetitive solid-state	document Livox `non_repetitive=True` convention
IMU (6-DoF)	`IMU`	✅ `accel_noise_density`, `gyro_noise_density`	—
IMU (9-DoF)	`IMU`	⚠️ no magnetometer noise field	put `mag_noise_density` in `metadata` until §4 gap M resolved
GNSS-INS	`GPS` + `IMU` `SensorBundle`	✅	bundle the two
6-axis F/T	`FORCE_TORQUE` + `n_axes=6`	✅	—
Vision tactile	`TACTILE_VISION`	✅	optionally fill `intrinsics` for the embedded camera
Array tactile	`TACTILE_ARRAY` + `tactile_grid`	✅	put magnetic / capacitive / barometric tech in `metadata`
Audio	`AUDIO`	✅	—
GPS	`GPS`	✅	—
Battery	`BATTERY`	✅	—

4. Identified schema gaps (proposals — none of these block any factory work above)

Each of the following is a proposal for a future ADR + minor SemVer bump of openral_core (CLAUDE.md §1.6). Until then, sensors using these technologies should encode the type via metadata["physical_modality"] and choose the closest existing modality.

Gap	Proposed enum addition	Why it matters	Closest current fit
E. Event / neuromorphic vision	`SensorModality.EVENT`	Prophesee EVK4, iniVation DAVIS — shipping in research drones and humanoids; very different data semantics (event streams, not frames)	misuse `RGB` today
T. Thermal / LWIR	`SensorModality.THERMAL`	Spot CAM+ thermal payload, FLIR Boson; common on inspection and SAR	misuse `IR` today
U. Ultrasonic / radar / proximity	`SensorModality.RANGE_1D` (covers ultrasonic + ToF range + radar)	Drone altitude, mobile-robot bumpers, humanoid foot proximity	misuse `DEPTH` today
B. Barometer / altimeter	`SensorModality.BAROMETER`	Drones, multi-floor mobile robots	metadata-only today
M. Magnetometer (standalone)	`SensorModality.MAGNETOMETER` (or extend IMU spec with magnetometer noise field)	9-DoF IMUs that publish magnetometer separately, indoor magnetic localization	inside `IMU` metadata today
C. Contact / binary touch	`SensorModality.CONTACT`	Foot contact switches on humanoids/quadrupeds, gripper-close detection	misuse `FORCE_TORQUE` (n_axes=1) or `TACTILE_ARRAY` (1×1) today
S. Stereo baseline field	extend `SensorSpec` with `baseline_m: float \\| None`	ZED, OAK-D, MultiSense — baseline is a first-class extrinsic, not metadata	inside `metadata` today
W. Wheel encoder / odometry	(no enum — keep as `nav_msgs/Odometry` topic)	Standard ROS 2 convention; bind via `WorldState.base_twist`	covered by `WorldState`

Recommendation: bundle E + T + U + B + M + C + S into a single additive ADR that extends SensorSpec in place against the pre-publish baseline (schema_version: "0.1"), with a doctest loading a fixture manifest.

5. Catalog & factory implementation status

The catalog ships only the sensor entries that are actually wired into a HAL adapter today. Each entry is addressable by stable id (<vendor>/<model>) through the SensorCatalog registry and surfaced via the openral sensor list / show CLI. Speculative entries (extra RealSense SKUs, Orbbec, Logitech C922 / Brio, Arducam, every 2D / 3D LiDAR, standalone IMUs, ATI / Bota / OnRobot F/T, GelSight DIGIT, XELA, AnySkin) were dropped in the cleanup pass; reintroduce them only when a robot manifest or HAL factory needs them.

What shipped (5 entries, 3 modules)

Module	Catalog ids	Modalities	Wired in
`realsense.py`	`intel/realsense_d435`, `intel/realsense_d435i`, `intel/realsense_d415`	RGB+DEPTH(+IMU) bundles	`ur10e_with_sensors`, `franka_panda_with_sensors`, `ur5e_with_sensors`
`usb_uvc.py`	`logitech/c920`	RGB	`so100_with_sensors`
`force_torque.py`	`robotiq/ft_300s`	FORCE_TORQUE (n_axes=6)	`ur5e_with_sensors`, `ur10e_with_sensors`

How to use

from openral_sensors import CATALOG

# List
print(CATALOG.list_ids())                       # 5 ids
print([e.id for e in CATALOG.filter(vendor="intel")])

# Build a SensorBundle / SensorSpec for a robot manifest
bundle = CATALOG.build("intel/realsense_d435i", name="head", parent_frame="head_link")
robot.sensor_bundles.append(bundle)

openral sensor list                                  # tabular catalog
openral sensor list --vendor intel
openral sensor list --modality rgb --json
openral sensor show intel/realsense_d435i --name head --parent-frame head_link

Not yet shipped — Tier 3 (gated on schema v0.3 ADR)

Module	Factories	Blocked-by
`stereolabs.py`	`zed_2i_bundle`, `zed_x_bundle`	gap S (stereo `baseline_m`)
`luxonis.py`	`oak_d_pro_bundle`, `oak_d_lite_bundle`	gap S
`event_camera.py`	`prophesee_evk4_spec`, `inivation_davis346_spec`	gap E
`thermal.py`	`flir_boson_spec`, `spot_thermal_payload_spec`	gap T
`proximity.py`	`maxbotix_ultrasonic_spec`, `teraranger_evo_spec`	gap U
`barometer.py`	`bmp388_spec`	gap B

6. Concrete style for new modules

Every new module follows the conventions used by realsense.py, usb_uvc.py, force_torque.py:

Module docstring with a runnable doctest example.
Nominal data-sheet constants at module top, behind comments citing the data-sheet revision (realsense.py).
One <model>_bundle() factory per multi-modal sensor returning a SensorBundle(sync="hardware" | "approximate", sync_tolerance_ms=…).
One <model>_spec() factory per single-modality sensor returning a SensorSpec.
SensorCatalog registration at the bottom of the module — every public factory is registered with replace=True so re-imports during tests are idempotent.
Optional bundle_to_node_params() and generate_launch_py() for ROS 2 wiring (mirrors realsense2_camera's composable-node pattern, RealSense only today).
Optional calibrate_*_cmd() for calibration helpers (chessboard for RGB, data-sheet preset for IMU, etc.).
Unit tests added to tests/unit/test_sensor_catalog.py covering: factory shape, modality composition, parent-frame propagation, encoding, range, metadata, catalog round-trip via model_dump_json / model_validate_json.
Re-export from python/sensors/src/openral_sensors/__init__.py (the side-effect import is what populates the catalog).

7. Summary — status & next steps

✅ Catalog pruned to actually-wired sensors. openral_sensors now ships 5 catalog entries across 3 vendor modules (realsense.py, usb_uvc.py, force_torque.py), SensorCatalog registry, openral sensor list / show CLI, full unit-test coverage. Any speculative entry (extra LiDAR, IMU, tactile, depth-camera SKUs, additional UVC cameras) was removed — reintroduce them on demand when a robot manifest or HAL adapter actually needs them. No schema change required.
Open one ADR for the §4 schema additions (event, thermal, range_1d, barometer, magnetometer, contact, stereo baseline) bundled together as schema v0.3, with a single migration entry and a regenerated docs/reference/schemas/.
After the ADR lands, ship Tier 3 modules (stereolabs.py, luxonis.py, event_camera.py, thermal.py, proximity.py, barometer.py) as separate small PRs, each with the test shape mandated in §6.
Optional follow-up: add a catalog: reference field on RobotDescription.sensors / RobotDescription.sensor_bundles so a robot.yaml can say - catalog: intel/realsense_d435i instead of inlining the whole SensorBundle. This is a tiny v0.2 → v0.3 schema migration; folding it into the same ADR keeps the migration count down.

Last reviewed: 2026-05-06. Re-review whenever a new humanoid platform ships or a major depth/lidar vendor changes status.