Live Training Visualization
Watch the training and validation loss curves update in real time as the network trains. See whether the network is converging, plateauing, or overfitting, all in your browser.
AI Neural Network Kinematics Simulator
このコンテンツはまだ日本語訳がありません。
Can a neural network learn to solve inverse kinematics? This simulator lets you find out by training a real TensorFlow.js model directly in your browser. Generate training data via motor babbling (random joint angles through forward kinematics), configure the network architecture, watch the loss curve drop in real time, then click anywhere on the workspace to compare the neural network prediction with the exact analytical solution. #NeuralNetworkIK #TensorFlowJS #MachineLearningRobotics
Open SimulatorWhat You Can Explore
Error Heatmap
After training, the workspace is color-coded by prediction error: green where the network learned well, red where it struggles. Errors are typically highest near workspace boundaries and singularities.
NN vs Analytical Comparison
Click anywhere and see both the neural network solution (purple) and analytical solution (green) simultaneously. Compare position accuracy and inference speed.
Architecture Experiments
Change the number of layers, neurons, activation function, dropout rate, and training data size. Build intuition for how network capacity and data volume affect kinematics learning.
Preset Configurations
| Preset | Samples | Layers | Neurons | Epochs | Purpose |
|---|---|---|---|---|---|
| Quick Demo | 1,000 | 1 | 32 | 50 | Fast overview, moderate accuracy |
| Standard | 5,000 | 2 | 64 | 100 | Good balance of speed and accuracy |
| High Accuracy | 20,000 | 3 | 128 | 200 | Best accuracy, longer training |
| Overfitting Demo | 200 | 3 | 128 | 300 | Demonstrates overfitting with small data |
Experiments to Try
- Dataset size effect: Train with 500 samples, note the mean error. Then train with 5,000 and 20,000. Plot how error decreases with more data (diminishing returns curve).
- Architecture search: Keep data at 5,000. Train with 1 layer/32 neurons, then 2/64, then 3/128. Save each as Experiment A to compare. More layers help, but with diminishing returns.
- Overfitting detection: Use the “Overfitting Demo” preset. Watch the training loss decrease while validation loss increases after some point. This gap is the hallmark of overfitting.
- Activation function comparison: Train identical architectures with ReLU, then sigmoid, then tanh. ReLU typically converges faster for this problem.
- Workspace error analysis: After training, enable the error heatmap. Notice that errors are highest near the workspace boundary (full extension) and near the base (full fold). These are where the IK mapping is most nonlinear.
- Speed comparison: Click around the workspace and compare inference times. The neural network produces answers in under 1 ms (one forward pass), while analytical methods are also fast for 2-DOF but scale differently for higher DOF.
How It Works
Training data generation: The network learns by example. Random joint angles are sampled uniformly, and forward kinematics computes the corresponding end-effector positions. This creates input-output pairs: (x, y) mapped to (q1, q2).
Network architecture: A feedforward neural network with configurable hidden layers and neurons. Input: normalized (x, y) position. Output: normalized (q1, q2) joint angles. Loss function: mean squared error.
The multi-solution problem: For most targets, multiple joint configurations reach the same position (elbow-up and elbow-down). A standard feedforward network learns an average, which may not correspond to any valid configuration. This is visible in the error heatmap as regions of higher error where both solutions exist.