Physics Prediction¶

Physics prediction component — Genesis-backed simulation with pure-Python fallback.

Cluster: Physical AI | Type: component | MCP Tools: None

Overview¶

Backend fidelity

Without genesis-world, Genesis-specific operations fail closed and the component uses pure-Python fallback dynamics. The fallback is useful for estimates and tests, not safety-critical physics prediction.

Physics simulation and prediction engine for articulated robots and rigid-body scenes. Implements forward dynamics, contact force estimation, material property lookup, trajectory generation, and stability scoring using a pure-Python backend by default, with optional Genesis solver backends (rigid, MPM, SPH, FEM) for high-fidelity simulation. Supports URDF robot loading and multi-solver scene generation when Genesis is available.

When to use:

Estimating joint trajectories and contact forces before simulated or human-reviewed physical planning
Assessing breakage probability or stack stability before a manipulation step
Generating interpolated waypoint trajectories for robot motion planning
Discovering backend availability before execution via source_capabilities and source_health

Fallback outputs use the canonical envelope: completion_state: qualified-draft, degraded: true, a warning_card, and top-level evidence. Physics prediction and simulation results (trajectories, stability, contact force, breakage, risk, generated scenes/trajectories, articulated/stepped state) are always completion_state: qualified-draft: a model output is a prediction, never self-certified as externally-validated ground truth (S-P0-1). Only non-prediction discovery/metadata/lifecycle ops (get_info, reset, source_capabilities, source_health, material get/set) use completion_state: verified. Failures/refusals are surfaced as blocked-escalated by response wrappers such as MCP. predict_outcome discloses its breakage model as linear_force_ratio with uncalibrated_heuristic confidence metadata.

Example:

from mvp.physics_prediction import PhysicsPredictionBlock, PhysicsPredictionInput

block = PhysicsPredictionBlock(name="physics")
result = block.infer(PhysicsPredictionInput(
    op="predict_trajectory",
    initial_state=[0.0, 0.0, 1.0, 0.0, 0.0, 0.0],
    forces=[[0.0, 0.0, -9.81]] * 20,
    steps=20,
    dt=0.05,
))
# result.value.predicted_trajectory → list of state vectors

Works well with: motor_control, mesh3d, sensory_fusion

Public API¶