Firefighter robot

Firefighter robot

Progress report. Last edited on 13 October 2026. The results date from 8 to 13 October 2026.

The firefighter robot walks to a small fire and puts it out with water. It uses the legs and the torso frame of the open-source Berkeley Humanoid Lite. A water gun or a small pump sits on top.

This page collects the work outside vision: the body, the payload, walking, the hardware, the aiming software and a test platform for the fire detector. Marc works on the fire detector, so this work stops at the interface to his detector.

Note: Nothing on this page is built or measured on the robot yet. Every result comes from simulation, CAD or research. Each section gives its status with one of these words: built, tested in simulation, measured on laptop, measured on board, planned.

Status

Work packageStatusMain resultFolder
Body conceptsplannedFive concepts on the same legs. Concept A, a water gun on a pan-tilt, is the suggested first build.cad/
Payload on the legstested in simulationThe worst leg torque for A with the real CAD parts is 5.2 Nm. The policy limit is 6 Nm.sim/payload/
Official walking policytested in simulationNo fall with any payload at 6 Nm. At 4 Nm, concept B with a full bottle falls in under 1 s.sim/sim2sim/
Closed looptested in simulationThe robot found the fire, walked, aimed and put it out from 1.9 m. It did not fall.sim/sim2sim/
Our walking policytested in simulationIt turns on the spot at 0.39 rad/s. It is slower and less robust than the official policy.sim/rl/
Pump inside the torsotested in simulationTwo printed tanks for a 33 g micropump: R1 inside the torso (0.63 L) and R2 on the sides (0.73 L). Both give the lowest centre of mass so far and walk like the bare robot. R1 is the suggested option.cad/torso_tank/, sim/payload/
Printable partsplanned25 parts with STEP and STL files: pan-tilt in 2 versions, adapters, splash guard, backpack and bench rig.cad/parts/, cad/backpack/, cad/bench/
Parts listplanned463,273 KRW plus Waveshare shipping. The lab limit is 500,000 KRW.docs/hardware/
Aiming code and driverstested in simulation45 tests pass. Nobody has run the code on real servos, a Pico or the robot.aim/, actuators/
Vision test platformtested in simulation270 synthetic samples, a camera-in-the-loop simulator and six test scenarios. All scenarios pass.sim/firescenes/, sim/sim2sim/

Body concepts planned

Five robot concepts side by side on the same legs
From left to right: A water gun turret, B backpack pump, C arm lance, D sealed shell, E tethered hose. Legs and torso are the real Berkeley meshes.
Concept A with the printed module (ST3215 version) and a stand-in gun.

All five concepts keep the Berkeley legs and torso frame. They differ in the water source, the aiming method and what sits on top of the hips. The Body concepts article describes each concept with renders.

ConceptWaterAimAdded massBuild effort
A Water gun turretGun tank, about 0.5 LPan-tilt, 2 servos~1.4 kgLow
B Backpack pumpTank and 12 V pump on the backHead pan-tilt, 2 servos~2.7 kgMedium
C Arm lance1 L tank and pumpRight arm, 5 motors~7.3 kgHigh
D Sealed shellAbout 1 L insideWhole body~2.3 kgMedium to high
E Tethered hoseExternal, no limitPan-tilt, 2 servos~1.0 kgLow to medium
Added masses are first estimates. For C, they include 5.0 kg of arms.

We suggest concept A first, because it needs the least new work. The gun keeps its own tank and pump, and a relay switches its trigger. The same deck and 40 mm hole pattern can later take the camera head of B or the valve of E. The pump of concept B is the upgrade for range and flow.

Payload on the legs tested in simulation

We built MuJoCo models of the robot with the payload of each concept. We then computed the joint torques in slow poses with inverse dynamics. One-leg stance gives the highest torque, and it happens in every step.

ModelPeak torque, one-leg stance
Biped, no payload4.3 Nm
A, first estimate5.9 Nm
A, real CAD parts5.2 Nm
B, first estimate3.6 Nm
B, real CAD parts, full bottle3.9 Nm
D6.0 Nm
E, 20 N hose pull5.2 Nm
C, with arms11.6 Nm
Pitch joints. The walking policy limit is 6 Nm. The motor model allows 20 Nm.
Heat maps of torque and push recovery for a 2 kg mass at different positions
A 2 kg mass moved over a grid: peak torque (left) and push recovery (right). Negative x is behind the torso.
Note: One-leg stance is a slow upper bound. The real gait sways and keeps each foot on the ground for a short time only.

Walking

Official policy with a payload tested in simulation

The Berkeley team ships trained walking policies. We ran policy_biped_25hz_a.onnx in our own MuJoCo loop with the payload models. Each test lasted 20 s.

Warning: The simulation has no motor model, no delay and no IMU error. The real robot will do worse.

Closed loop: see, walk, aim, spray tested in simulation

Official policy and our aiming code. Concept A with the real parts, fire behind and to the right.

We joined the official policy at 6 Nm, our aiming code and concept A with the real CAD parts in one MuJoCo run.

Robot path and controller states during the closed-loop run
Robot path (left) and controller states (right) for the same run.

Torque limit and motor heat tested in simulation

Quantity, walking at 6 Nm, A and B with real partsValue
Highest RMS joint torque3.8 Nm
Copper loss per motor (estimate from the Berkeley resistance and torque constant)0.4–2.4 W
Time a joint is above 4 Nm47 %
Warning: No source gives a continuous torque or a thermal limit for the Berkeley actuator. The heat in the printed gearbox is unknown. Before the robot walks with a payload, run a 2-minute walking test at 6 Nm and log the motor temperatures.

Our policy tested in simulation

We trained our own policy with MuJoCo Playground on the RTX 3090. During training, a payload of 0 to 2.5 kg sits on the torso. Each run took 16 to 37 min.

Training setup0 kg1.5 kg3.5 kg
6 Nm, payload 0–2.5 kg29 %42 %59 %
6 Nm, payload 0–3.5 kg, knee ≤ 1.4 rad43 %48 %61 %
20 Nm, payload 0–2.5 kg1 %3 %10 %
Episodes that end in a fall with random pushes, by test payload.
Note: The Berkeley action scale (0.25) gave a policy that only stood. We used the Playground value (0.5).

A policy that turns on the spot tested in simulation

Our new policy turns on the spot. Concept A with the real parts, 6 Nm.

We trained a second policy with the same interface as the official policy: 45 inputs, 12 outputs, 25 Hz. Its ONNX file can replace theirs. We tested both with the same harness, with the training noise on.

Concept A, real parts, 6 NmOursOfficial
Walk, 0.25 m/s asked0.15 m/s0.23 m/s
Turn on the spot, 0.5 rad/s asked0.39 rad/s0.04 rad/s
Walk and turn (m/s, rad/s)0.17, 0.440.23, 0.48
Falls with pushes, of 108–95–7
Concept B, 2.8 kgfallswalks

Our policy turns on the spot, which the official policy cannot do. But it walks slower, falls more under pushes and falls with concept B. It is not ready. The next step is a longer training with pushes and heavier payloads.

Hardware

Pan-tilt module planned

Adapters, splash guard and backpack planned

Warning: With the backpack, set a software knee limit of 1.4 rad. Otherwise the lower leg can touch the backpack.

Bench rig planned

Bench rig, one full turn.

The bench rig lets us test the module, the water and the aiming before the robot walks.

The test plan is in cad/bench/README.md:

  1. Calibrate the servos.
  2. Measure the jet speed.
  3. Align the camera with the nozzle.
  4. Aim at paper targets.
  5. Put out a small fire outdoors.

Parts list planned

The parts list covers the water system, the aiming and the safety parts. The lab pays. 3D-printed parts cost nothing, because the lab has printers. The lab already has the motors and all other parts of the Berkeley robot. The cameras belong to the vision budget.

The total is 463,273 KRW with VAT, plus the Waveshare shipping. The lab limit is 500,000 KRW.

GroupKRW
Water165,156
Aiming126,776
Electronics103,740
Safety44,671
Wiring22,930
Total463,273
GroupItemQtyTotal (KRW)ShopPrice
WaterSeaflo SFDP1-010-040-21 12 V diaphragm pump (3.8 L/min, 0.28 MPa), 1 + 1 spare272,600Devicemartprice of 8 Oct
Nalgene 1 L wide-mouth bottle (fits the printed cradle)127,000Kolon Mallchecked 9 Oct
Braided PVC hose, 8 mm inner diameter, per metre33,300Devicemartprice of 8 Oct
Brass fitting, PT 1/4 male to 8 mm barb (hose to nozzle)23,718NaviMallchecked 9 Oct
Brass coupling, PT 1/4 female-female11,529NaviMallchecked 9 Oct
Hose clamp, 8–12 mm61,980NaviMallchecked 9 Oct
Sanga SJN02 solid-stream nozzle, PT 1/4, 2 mm orifice (pack of 10)121,989NaviMallchecked 9 Oct
Daiso electric water gun (about 100 mL, first test)15,000Daiso Mallchecked 9 Oct
Rechargeable electric water gun, 400 mL (demonstration)128,04011stchecked 9 Oct
AimingWaveshare ST3215 12 V serial bus servo (30 kg·cm), pan and tilt258,212WaveshareUSD × 1,400
Waveshare Bus Servo Adapter (A), USB to servo bus16,986WaveshareUSD × 1,400
6808 bearing (40 × 52 × 7 mm), pan axis117,589NaviMallchecked 9 Oct
M5 T-nuts for 2020 profile (pack of 100)125,289NaviMallchecked 9 Oct
M3 × 4 mm heat-set inserts (pack of 50)111,880ICbanqchecked 9 Oct
M4 brass inserts206,820NaviMallchecked 9 Oct
ElectronicsDC-DC converter, 20–96 V to 12 V, 10 A, waterproof131,790ICbanqchecked 9 Oct
Raspberry Pi Pico 2 (relay control)112,100ICbanqchecked 9 Oct
2-channel 5 V relay module (channel 1: gun trigger, channel 2: pump)11,760ICbanqchecked 9 Oct
SB560 Schottky diode (pump flyback)51,870ICbanqchecked 9 Oct
NEXT UH305 powered USB 3.0 hub126,30011stchecked 9 Oct
Adafruit 3626 DC power meter (100 V, 20 A), to measure the current129,920ICbanqchecked 9 Oct
Safety22 mm latching e-stop, 2 NO + 2 NC (KSEP2260-22)116,280ICbanqchecked 9 Oct
Power relay, 24 V coil, 30 A11,760ICbanqchecked 9 Oct
Inline blade fuse holder1792ICbanqchecked 9 Oct
Blade fuse, 10 A52,475ICbanqchecked 9 Oct
Blade fuse, 5 A52,475ICbanqchecked 9 Oct
ABC powder extinguisher, 3.3 kg (for the tests)120,889NaviMallchecked 9 Oct
WiringXT30 connector pair (tap on the robot 24 V bus)56,600ICbanqchecked 9 Oct
XT60 connector pair26,380ICbanqchecked 9 Oct
Silicone cable 16 AWG, per metre (2 m red + 2 m black), plus 2,800 KRW shipping45,00011stchecked 9 Oct
DC jack 5.5 × 2.1 mm with cable21,100Elepartschecked 9 Oct
Dupont wires (65 pieces)11,320ICbanqchecked 9 Oct
Heat-shrink tubing (100 pieces)12,530ICbanqchecked 9 Oct
Optional, not in the total22 AWG UL1007 wire kit123,100Elepartschecked 9 Oct
Third ST3215 servo as a spare129,106WaveshareUSD × 1,400
ST3215 from a Korean shop (if the order from abroad is a problem)174,789NaviMallchecked 9 Oct
KITZ 3/8" Y-strainer for the pump inlet (heavy)145,089NaviMallchecked 9 Oct
Note: Prices change. Most prices were checked on 9 October 2026. The Devicemart site refused connections on that day, so its prices date from 8 October. The Waveshare prices are USD prices × 1,400. The original list, in French, is docs/hardware/bom_extinction.csv.

Power and wiring planned

Wiring diagram for concept A
Wiring for concept A. Open the image for full size.

Power comes from the robot's 24 V bus through a 10 A fuse, an e-stop relay and a 24 V to 12 V converter. The e-stop cuts all add-on power. It does not stop the leg motors.

# Pico 2 pins
GP14 -> relay IN1   water-gun trigger
GP15 -> relay IN2   12 V pump
GP16 <- e-stop NC   open = stop

# Servo ids
1 = pan, 2 = tilt

Software

Overview tested in simulation

cameras --images--> fire detector (Marc)
                         |
                         |  UDP 5600: detections
                         v
robot pose --UDP 5605--> aim/run.py --UDP 5601: pan, tilt, water--> actuator_node.py --> servos, pump, gun
                         |    ^                                         |
                         |    +------- UDP 5603: measured angles -------+
                         |
                         +--UDP 5602: vx, yaw rate--> walk_bridge.py --> Berkeley walking policy

Every message is one JSON line over UDP on the robot computer. The detector sends only pixel positions and boxes. The aiming code computes the fire position, the jet angles and the walk command.

Detector interface

Version 1, the current interface: the detector sends one JSON line for each fire and frame to 127.0.0.1:5600.

{"timestamp": 12345.678, "camera": "rgb", "u": 331.5, "v": 298.0,
 "bbox": [310, 280, 352, 316], "confidence": 0.87,
 "max_temp_c": null, "depth_m": null}

The timestamp is time.monotonic() on the robot computer at the time of the image.

Version 2, a proposal: the detector sends one frame message for each camera image, also when it sees no fire. An empty list means that the image has no fire.

{"timestamp": 12345.678, "camera": "rgb", "frame": true,
 "detections": [{"u": 331.5, "v": 298.0, "bbox": [310, 280, 352, 316], "confidence": 0.87}]}

With version 1, aim/run.py cannot tell "no fire in this image" from "the image is late". The software-in-the-loop test found 9 to 24 % of controller ticks without a new image. aim/run.py accepts both versions.

Questions for Marc:

Aiming code tested in simulation

Top view of a toy simulation: the robot turns, walks and sprays
Toy simulation of concept A.

The aiming code starts where the detector stops. It finds the fire on the floor, walks closer, aims the jet and decides when to spray.

Tip: You can test without a camera. aim/send_test_detections.py sends fake detections to aim/run.py.

Actuator drivers tested in simulation

Terminal output of the self-test in mock mode
Self-test in mock mode. The same 7 checks run on the first day with the real parts.
st3215.py
Driver for the Feetech STS servos. The tests use the packet examples from the Feetech manual.
pico_fw/main.py
Relay firmware for the Pico 2. The gun and the pump switch off after 0.5 s without a message, or when the e-stop opens.
actuator_node.py
Sends the aim commands to the servos and relays. Sends the measured angles back on UDP 5603. A mock mode runs without hardware.
walk_bridge.py
Replaces the Berkeley gamepad input. Sends a zero command after 0.5 s without a message.
selftest.py
7 checks for the first day with the parts: ports, servos, relays, watchdog, e-stop and UDP. It asks before every move.

The full chain runs in mock mode: test detections, aiming, actuators, then done after 3.3 s of water. 10 tests pass.

Note: Nobody has run the drivers on real servos, a Pico or the robot yet.

Test platform for vision

We have no robot and no bench to test the fire detector on yet. These tools let Marc test the detector with the aiming code before the hardware exists.

Camera views tested in simulation

Camera views before and after the pose change
Camera views before and after the recommended pose change, 90° lens.

We rendered the view of each robot camera in a test room with fires at 1, 2, 3 and 5 m. We also computed how many pixels a 20 cm fire covers for each sensor.

Sensor1 m3 m5 m8 m
MLX90640, 55°, 32 × 246 × 102 × 31 × 21 × 1
Lepton 3.5, 160 × 12030 × 3710 × 126 × 74 × 5
256 × 192 USB thermal48 × 6316 × 2110 × 136 × 8
RGB 1280 × 720, 90°128 × 16043 × 5326 × 3216 × 20
Size of a 20 cm fire in pixels (width × height).

Synthetic fire scenes tested in simulation

A 10-second walk toward a fire: RGB (left) and thermal-like image (right).
Grid of synthetic RGB and thermal image pairs with labels
Synthetic samples: RGB and thermal-like pairs with their labels.
Warning: This is not a real sensor model. It has no noise and no blur. Use it to test the pipeline, not to train the final detector.

To make more samples, for example numbers 250 to 749:

blender -b --factory-startup -P sim/firescenes/firescenes.py -- samples 250 750

Camera in the loop tested in simulation

The gun turns to the fire. Top right: the robot camera with the detection box.

Every 100 ms the simulator renders the gun camera and calls detect(image, t). The detector output goes to the aiming code, and the robot walks and sprays in MuJoCo.

To plug in a detector:

  1. Write a Python file with a function detect(rgb, t) that returns a list of FireDetection (aim/detection.py).
  2. Run the scenario suite with --detector and the path of that file.
MUJOCO_GL=egl sim/sim2sim/.venv/bin/python sim/sim2sim/camera_loop.py --quick --detector path/to/my_detector.py
from detection import FireDetection  # camera_loop.py puts aim/ on the import path


def detect(rgb, t):
    """rgb: 480 x 640 x 3 uint8 RGB image of camera "rgb". t: time of the image, s."""
    boxes = my_model(rgb)  # list of (u_min, v_min, u_max, v_max, score)
    return [FireDetection(t, "rgb", (u0 + u1) / 2, v1, (u0, v0, u1, v1), score)
            for u0, v0, u1, v1, score in boxes]
Note: The image is RGB, not BGR. OpenCV code must convert it with cv2.cvtColor(rgb, cv2.COLOR_RGB2BGR). The camera intrinsics are fx = fy = 320 px, cx = 320 px and cy = 240 px.

Test scenarios tested in simulation

Scenario "two fires": the robot puts out both.
ScenarioResultFire out at
Fire ahead, 2.5 mpass7.3 s
Fire behind and to the rightpass18.9 s
Two firespass10.0 s and 38.3 s
Lamp only, no firepassno spray
Fire 0.5 m in frontpass7.0 s, one spray
Fire 6 m awaypass38.7 s
Baseline colour detector, concept A with the real parts.

Software in the loop tested in simulation

In the software-in-the-loop test, the real programs run as separate processes and talk over UDP, as on the robot. The simulator replaces only the camera, the legs and the servos.

MUJOCO_GL=egl sim/sim2sim/.venv/bin/python sim/sim2sim/camera_loop.py --sil --repeat 5
Warning: All programs ran on one computer. A detector on another computer has a different clock, and aim/run.py then rejects every message. This case is not tested.

Decisions for the team

IDTopicQuestion
D1ConceptA first, with B as the upgrade on the same module?
D2Servo typeST3215 bus servo (feedback, 12 V) or DS3218 hobby servo? Both CAD versions exist.
D3Torque limitRun the robot at 6 Nm instead of the 4 Nm default? Check the motor heat first.
D4OrderWho sends the parts list to the lab, and when?
D5ComputerDetector and aiming code on the robot's N95 PC, or on a laptop over Wi-Fi?
D6TankFor the pump inside the torso: tank R1 inside the torso or R2 on the sides? See Pump inside the torso.

Next steps

  1. Order. Send the parts list. Print the module and the bench rig.
  2. Assemble. Wire the servos, the Pico, the relay and the e-stop. Flash the Pico.
  3. Calibrate. Set the servo zeros, measure the jet speed and align the camera with the nozzle.
  4. Test on the bench. Run the detector, the aiming code and the spray on paper targets. Then put out a small fire outdoors.
  5. Robot. Run the policy at 6 Nm with the payload. Do a first standing test.
Warning: Have the e-stop and the extinguisher ready before the first test with water or fire.

Files

All folders are in the main branch of the vision-sobang repository. Each folder has a README with the method, the commands and the limits.

ContentFolder
Concept renders and Blender scriptscad/, cad/renders/
Printable parts (STEP, STL)cad/parts/, cad/backpack/, cad/bench/
Payload and policy simulationssim/payload/, sim/sim2sim/, sim/rl/
Aiming code and driversaim/, actuators/
Vision test platformcad/camera_views/, sim/firescenes/
Parts list, wiring, Berkeley deployment notesdocs/hardware/, docs/locomotion/