Skip to content

Perception Audit: 312-scene CoC analysis — findings submitted to team #47

Description

@dennisdarrow

I've completed a large-scale perception audit of Alpamayo 1 10B across 312 NuRec intersection scenes (61,150 timesteps). The full 23-page report has been shared privately with the team.

Objective: Determine whether the model’s perception claims are geometrically consistent with what the cameras can actually observe.

Methodology:

  • Colocated AlpaSim deployment on Lambda Labs 4×H100 SXM5 (~$442 total compute)
  • 4-camera configuration at 1080p/10Hz: front_tele, front_wide, cross_left, cross_right
  • Camera-aware visibility model with angular occlusion culling and per-type motion thresholds
  • Audit script extracts two data streams per scene (ASL Telemetry Ground Truth and CoC Text) and aligns them at 10Hz
  • 7 verdicts: Agreement, Dynamic Hallucination, Phantom Object, Geometric Hallucination, Potential Under report, Planning Only, Unclassified.
  • Bumper proximity scoring (Critical <2m, Warning <5m, Info ≥5m from front bumper)
  • Range sensitivity validated across ±50% camera range variation (±0.4pp spread)

Key findings have been shared privately with the team. Happy to share more details once the team has had a chance to review.
This work reflects independent research conducted in a personal capacity.

Dennis Darrow

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions