Cracking Mobile Manipulation: From Failure Diagnosis to Agentic Robot System

Key points

A pretrained subtask VLA achieves 80% success on atomic tasks but only 19% when the skills are composed into mobile manipulation.
Analysis of 100 episodes shows the primary failures split into task planning 21.0%, navigation and base positioning 25.9%, and local manipulation 53.1%.
Keeping the policy fixed and adding demonstration-based planning guidance, adaptive visual memory, accessibility-aware base adjustment, and end-effector adjustment raises RoboCasa365 success from 19% to 47%.
On held-out episode seeds the finalized system reaches 49% against the 18% baseline, with a paired difference significant at p < 0.001.
On BEHAVIOR-1K performance improves from 14% to 37%, and on two real-robot tasks from 5/20 to 13/20.
As interventions are added, the remaining failures shift toward manipulation: 47 of the 53 full-system failures are manipulation failures.

External research summaries. These are not HDATF publications or measured product results.