skip to content
← back

DSO National Laboratories · May – Jul 2026

Meta-Robots

Agentic Improvement for Embodied AI

A demo of one episode. A navigation demonstration from the HM3D ObjectNav setting: first-person RGB, semantic segmentation, depth, and top-down maps as the agent searches for a couch. The episode succeeded (88 steps, SPL 0.86). A single episode illustrates the system; it does not by itself establish overall performance. The aggregate benchmark result is reported separately below.

overview

Meta-Robots extends Meta-Harness from language-based systems to embodied ObjectNav. It asks how an agentic system can diagnose navigation failures, hypothesise their causes, implement interventions, and evaluate the effect.

The work lives in simulated 3D environments and touches perception, planning, and navigation. It improves the surrounding system (the harness); it does not retrain model weights or introduce a new trained world model.

main contributions

  1. 01

    Agentic improvement framework

    A research and evaluation loop for proposing and testing modifications to an embodied navigation system.

  2. 02

    Two-phase proposal architecture

    Structured hypothesis generation is separated from constrained implementation.

  3. 03

    Failure analysis

    Trajectory telemetry, behavioural fingerprinting, cross-scene failure analysis, and Qwen2.5-VL visual criticism grounded in execution logs.

  4. 04

    ObjectNav evaluation

    Built around the ASCENT baseline in Habitat / HM3D.

results

baseline success · ASCENT
65.4%
final success · HM3D ObjectNav
68.9%
absolute improvement
+3.5 pp
relative improvement
≈5.4%

Absolute change is in percentage points; the relative figure is the same change as a share of the baseline. The precise HM3D evaluation split and conditions are pending confirmation.