Four groups of three joints.
- Rear left
- 0 · 1 · 2
- Rear right
- 3 · 4 · 5
- Front left
- 6 · 7 · 8
- Front right
- 9 · 10 · 11
Each group follows hip → thigh → knee. Calibration offsets follow the installed servo and horn, rather than the socket number.
Robotics & mechatronics · Research project
SpotMicro.
Pitch and roll
stabilization.
I integrated the robot’s electronics and developed a controller that adjusts all four legs to correct pitch, roll, and combined diagonal tilt.
Hardware-tested leveling proof of concept
- Controller
- Teensy 4.0
- Feedback
- MPU6050 accelerometer + gyro
- Control loop
- 50 Hz nominal
- Current milestone
- Pitch + roll leveling
01 / Integration
Make each joint predictable before closing the loop.
The starting point was a Nova SM3-derived quadruped. My work focused on electrical integration, servo calibration, sensor orientation, and developing and testing the leveling controller.
A rewiring change moved two leg groups to different driver channels. I reordered the calibration offsets with the physical servos, then used a sequential movement test to confirm each joint before running coordinated motion.
The robot uses eight 20 kg-class hip/thigh servos, four 35 kg-class knee servos, and separate regulated logic and servo rails from a 3S LiPo. The PCA9685 drives the servos at 60 Hz.
Check identity, direction, then response.
- InspectCheck power rails, common ground, and wiring with power off.
- ExerciseTest channels 0–11 individually: 90° → 102° → 78° → 90°.
- CalibrateApply per-servo offsets and confirm mirrored left/right motion.
- MeasureVerify the IMU’s axes before enabling pitch and roll correction.
02 / Attitude estimation
Separate tilt from the robot’s own motion.
Accelerometer-only control reacted to vibration and sudden movement. The current controller combines short-term gyro motion with an accelerometer reference, reducing that reference’s influence when the readings disagree.
MPU6050 → Teensy 4.0
Adaptive complementary filter
Installed axes:
+X rear · +Y right · +Z up
- 01
Calibrate at rest
Average 250 samples after entering the stance. Record pitch/roll targets and gyro biases, including the mounting offset.
- 02
Predict with the gyro
Integrate angular rate using the measured time between control updates.
- 03
Check accelerometer trust
Compare acceleration magnitude with gravity and compare the accelerometer angles with the gyro prediction.
- 04
Correct the estimate
Blend the filtered accelerometer reference with a nominal weight of 0.035 × trust. Filter the resulting control error.
Implementation: estimation and controller architecture ↗
03 / Four-leg control
Two axis commands.
Four coordinated corrections.
Pitch and roll each use proportional correction plus gyro-rate feedback. A four-corner mixer combines them, then scales the corrections together when a diagonal would exceed the available travel.
When one side of the board rises, the robot shortens the legs on that side and extends the opposite legs. Enable JavaScript to try the interactive illustration.
PD correction
Pitch: Kp 2.2 · Kd 0.06
Roll: Kp 1.8 · Kd 0.06
Deadbands reduce small corrections: 0.40° for angle error and 1.20°/s for angular rate.
Mirrored joints
Rear: +pitch ± roll
Front: −pitch ± roll
Left thighs/knees use 90° + c / 90° − c; right joints mirror those signs. Hips hold their calibrated center.
Limit and smooth
25° base compression · ±20° correction
60°/s maximum correction slew rate
Apply channel offsets, constrain joint commands, and retain fractional angles until pulse conversion.
04 / Physical testing
What worked, and where it stopped.
Physical tests demonstrated correction in pitch, roll, and combined diagonal tilt, with recovery as the support surface returned toward level.
Two-axis leveling
- Pitch correction in both directions
- Roll correction in both directions
- Combined diagonal correction
- Recovery toward the calibrated stance
These are qualitative hardware observations. Angle/frequency logs, repetition counts, and measured settling times are not yet available.
High-frequency shaking
During a progressively faster table-shaking test, the robot overreacted and tipped over even though the tilt amplitude was modest.
The controller reduces accelerometer trust during disturbances, but it does not yet recognize sustained vibration or latch a fall state. A safe disturbance-frequency limit has not been measured.
35° is a software threshold, not a demonstrated incline capability.
Above that deviation from the startup target, desired corrections return toward the base crouch with slew limiting. Servos remain powered, and correction can resume when the estimate returns within the threshold.
Development archive
The earlier pitch-control test.
This recording documents the original v1 pitch milestone. The current v3 controller adds roll, diagonal mixing, and adaptive fusion; a matching v3 recording is not yet in the repository.
Read the test record05 / Next development
From body leveling to controlled foot placement.
The next stage is to measure the leg geometry, establish single-leg inverse kinematics, and build a slow crawl gait. Quantitative disturbance logging and explicit vibration/fall handling are also needed.
This work uses a Nova SM3-derived platform. My contribution covers electromechanical integration, calibration, control development, and physical testing; the mechanical platform and upstream reference firmware are credited in the repository.