KiRecommendedSpatial

Robotics & Autonomous Systems

The Problem

The planned path violates joint limits that exist outside the token sequence. Trajectories assume instantaneous velocity changes because inertia is continuous physics that the model treats as discrete decisions. A loaded arm at full extension cannot reverse direction instantaneously, but "reverse direction" is a valid token.

How Ejentum Solves It

One API call forces your model to validate collision geometry, inertial limits, and boundary integrity at every planning step before the trajectory is finalized.

The Failures

  • 01

    The Pattern

    Planned paths violate boundary integrity: walls, joint limits, and clearance constraints not enforced

    Why It Happens

    The model plans in token space where "move left 30 degrees" is always a valid sequence. Physical boundaries are external constraints that exist outside the generation process.

    The Resolution

    SP-001Topology Validator

    Validates every path segment against topological constraints: walls are impenetrable, floors support weight, and containment hierarchies are respected.

  • 02

    The Pattern

    Trajectories assume instantaneous velocity changes, ignoring inertia and deceleration requirements under load

    Why It Happens

    Velocity changes are discrete decisions in the plan. Inertia and deceleration are continuous physics that must be computed, not predicted. The model generates the desired state without computing whether the transition is physically achievable.

    The Resolution

    SP-018Kinetic Momentum Tracker

    Models inertia, momentum, and force constraints for every movement. A robotic arm at full extension cannot reverse direction instantaneously.

  • 03

    The Pattern

    Static obstacles accounted for but dynamic obstacles (moving humans, other robots, conveyor items) not predicted

    Why It Happens

    Static obstacle maps are part of the input context. Dynamic obstacles require future-state prediction over a time horizon, which is a simulation task the planning model does not perform.

    The Resolution

    SP-030Dynamic Scene Projector

    Projects the future positions of all dynamic entities in the scene, incorporating predicted trajectories into the planning constraint set.

The Evidence

+7.4pp on spatial tasks

EjBench, 30 spatial tasks

Motion planning has strict physical constraints with single correct feasibility answers. A single scaffold that enforces intermediate state validation outperforms multi-ability injection. Spatial domain: Ki +7.4pp, Haki regressed -2.8pp.

EXT-SP-200.1430.857 Haki

Task tracked physical object positions through multiple state changes and perspective shifts. Baseline confused what the observer knew with where the object actually was. Haki enforced reference stability: each belief state anchored to direct observation only.

Run your next motion plan through the API. See how the scaffold catches the physics violation your planner generated as a valid token sequence.