Bipedal IMU Fall-Risk Telemetry (2,400 labeled windows)
A labeled IMU telemetry dataset for training balance/fall-risk classifiers on bipedal and humanoid robots. 2,400 rows, each a 0.5 s summarized window with 9 physically-motivated features: accel x/y/z (g), gyro x/y/z (deg/s), roll, pitch (deg), and angular-velocity magnitude (deg/s), plus a binary label_fall_risk. Balanced 50/50 between stable stances and tip/tumble events generated from a gravity-referenced motion model. CSV with header; drop-in for scikit-learn, PyTorch, or TensorFlow. This is the exact dataset used to train the companion TipGuard ONNX model.
Demonstration clipfree preview
Left: real-world source footage. Right: a signal derived from it (inter-frame motion field or edge/feature map). Source & license are credited in the clip.
Preview
Free preview — no purchase requiredFrameworks
- PyTorchONNX Runtime
Hardware
- Not specified
Sensors
- Not specified
Specs
- Size: 1.77 MB
- Duration: 10s
Tags
Free
MIT license
bipedal-imu-fall-risk-telemetry-2-400-labeled-windows.zip · 1.77 MB
Casper White✓
Trust score
Would you deploy this in production?
Not enough votes yet. (0/3 so far)
Sign in to vote.
Reviews
0 reviews
No reviews yet. Be the first to share your experience.
More from Casper White
View profile →TipGuard — Bipedal Fall-Risk Detector (ONNX)
TipGuard is a lightweight (4.7 KB) ONNX classifier that flags imminent loss-of-balance on bipedal/humanoid robots from a single IMU window. Input is a 9-feature vector — accel x/y/z (g), gyro x/y/z (deg/s), roll, pitch (deg), and angular-velocity magnitude (deg/s); output is a 2-class softmax probability [stable, fall_risk]. Feature normalization is baked into the graph, so you feed raw IMU readings directly. Trained on the companion Bipedal IMU Fall-Risk Telemetry dataset (2,400 labeled 0.5 s windows) and exported at opset 17. Runs in microseconds on Jetson Orin / any ONNX Runtime target — ideal as a safety reflex node. Validated example: a stable stance returns P(fall)=0.00; a 50+ deg/s tumbling window returns P(fall)=1.00.
DiffDrive-Mini — Differential-Drive Robot URDF
A clean, simulation-ready URDF for a compact differential-drive mobile robot. Includes base_link, two continuous-axis drive wheels, a spherical caster, and a top-mounted LiDAR link — each with mass and inertia tensors set for stable physics in Gazebo / Isaac Sim. Wheel separation 0.21 m, wheel radius 0.05 m, so it drops straight into a ros2_control diff_drive_controller or nav2 stack for SLAM and navigation experiments. Minimal, dependency-free, and easy to extend with cameras or arms.
Related assets
Bipedal IMU Fall-Risk Telemetry (2,400 labeled windows)
A labeled IMU telemetry dataset for training balance/fall-risk classifiers on bipedal and humanoid robots. 2,400 rows, each a 0.5 s summarized window with 9 physically-motivated features: accel x/y/z (g), gyro x/y/z (deg/s), roll, pitch (deg), and angular-velocity magnitude (deg/s), plus a binary label_fall_risk. Balanced 50/50 between stable stances and tip/tumble events generated from a gravity-referenced motion model. CSV with header; drop-in for scikit-learn, PyTorch, or TensorFlow. This is the exact dataset used to train the companion TipGuard ONNX model.
Humanoid Pick-and-Place Telemetry (110 frames, RGB camera)
Pick-and-Place for the Humanoid — 6877 frames of synchronized joint state at 50 Hz. Captured/authored with RGB camera in the loop and validated against real Humanoid kinematics. Drop-in ready for training, sim-to-real transfer, or on-robot deployment.
TipGuard — Bipedal Fall-Risk Detector (ONNX)
TipGuard is a lightweight (4.7 KB) ONNX classifier that flags imminent loss-of-balance on bipedal/humanoid robots from a single IMU window. Input is a 9-feature vector — accel x/y/z (g), gyro x/y/z (deg/s), roll, pitch (deg), and angular-velocity magnitude (deg/s); output is a 2-class softmax probability [stable, fall_risk]. Feature normalization is baked into the graph, so you feed raw IMU readings directly. Trained on the companion Bipedal IMU Fall-Risk Telemetry dataset (2,400 labeled 0.5 s windows) and exported at opset 17. Runs in microseconds on Jetson Orin / any ONNX Runtime target — ideal as a safety reflex node. Validated example: a stable stance returns P(fall)=0.00; a 50+ deg/s tumbling window returns P(fall)=1.00.