Business Context

RoboPick, a warehouse robotics company operating 8,000 pick-and-place arms across fulfillment centers, needs a fast inverse kinematics model to convert target end-effector poses into joint angles. The controls team wants an ML-based regressor that can approximate valid joint configurations quickly enough for motion planning warm starts.

Dataset

The training data was generated from a calibrated 6-DOF robotic arm simulator and validated against production telemetry. Each row represents one reachable target pose and one valid joint-angle solution.

• Size: 420K samples, 16 input features, 6 regression targets
• Target: Continuous — 6 joint angles in radians for a valid arm configuration
• Class balance: Not applicable; regression problem with multimodal target behavior near singularities
• Missing data: ~2% missing in payload metadata, <1% missing in previous joint state from dropped telemetry

Success Criteria

A good solution should achieve mean absolute joint error < 0.06 radians, 95th percentile Cartesian position error < 2.5 cm after forward kinematics validation, and stable performance on edge-of-workspace poses.

Constraints

• Inference latency must stay below 10 ms per pose on CPU.
• The model should be simple enough to retrain weekly and debug when failures occur near singularities.
• Predictions must respect joint limits after post-processing.

Deliverables

1. Build a supervised regression pipeline for inverse kinematics.
2. Explain what inverse kinematics is and why this is a multi-output regression problem.
3. Design preprocessing and feature engineering for pose and robot-state inputs.
4. Train and evaluate at least one strong baseline and one production-ready model.
5. Report metrics in both joint space and task space, and describe failure modes near singular configurations.