Technical Deep Dive

Physics-Based World Modeling

FieldSpace runs a live physics model of the scene beside your neural networks, validating perception and pruning impossible trajectories in real time.

Neural Nets See Patterns, FieldSpace Enforces Physics

Your perception models detect and classify. FieldSpace validates those detections against kinematic and geometric constraints, then forecasts feasible futures.

What Your Neural Networks Do

  • Pattern Recognition

    Detect cars, pedestrians, lanes, signs from camera/LiDAR

  • Classification

    Label objects by type, pose, and semantic category

  • Scene Understanding

    Identify road structure, free space, static obstacles

Keep your existing perception models—no retraining required.

What FieldSpace Adds

  • Physics Validation

    Reject detections that violate kinematic constraints (e.g., car at 200mph)

  • Trajectory Forecasting

    Short-horizon rollouts of feasible futures for each object

  • Safety Envelopes

    Hard constraints on ego motion (speed, accel, collision)

Integrates via ROS 1/2, gRPC—outputs constrained trajectories to your controller.

How the Physics Engine Works

FieldSpace maintains a live world model using partial differential equations (PDEs) to represent motion, contact, and geometry constraints.

1

Perception Fusion

Ingest detections from your perception stack (bounding boxes, point clouds, segmentation masks). Fuse camera, LiDAR, radar, IMU into a consistent state.

  • • Multi-modal sensor fusion
  • • Kalman filter tracking
  • • Coordinate frame alignment
2

Physics Model

Maintain a live PDE-based world model that enforces motion and geometry constraints. Each object has a velocity field and feasibility envelope.

  • • Velocity field (traffic flow)
  • • Kinematic constraints (max accel)
  • • Geometry collision checks
3

Trajectory Validation

Short-horizon rollouts forecast feasible futures. Prune trajectories that violate physics or safety. Output safe plans to your controller.

  • • Short-horizon prediction (0.5-3s)
  • • Reject infeasible motion
  • • Constrained trajectory output

Under the Hood: Modulus Calculation Core

FieldSpace is powered by Modulus, our deterministic solver for kinematics, contact, and geometry constraints.

What is Modulus?

Modulus is a PDE solver optimized for real-time autonomy. It maintains a physics-grounded world model, runs short-horizon forecasts, and rejects impossible motion before it touches your controller.

  • Deterministic: Same inputs = same outputs, every time
  • Fast: 1,974 FPS for traffic field computation on GPU
  • Scalable: Handles 50+ objects in real time
  • Portable: Runs on CPU, GPU, edge accelerators

Key Performance Metrics

Traffic Field Update Rate1,974 FPS
Decision Latency<5ms
Objects Tracked50+
Prediction Horizon0.5-3s

See the Physics Engine in Action

Run FieldSpace on your logs in Week 1 of a pilot. Measure baseline KPIs and see how physics validation improves your stack.

4-week pilot. KPI-driven. Integrates with ROS 1/2 and gRPC.