What they did

The authors built 2AM around a strict division of labor. A multimodal agent retains task history and converts it into subtask instructions plus optional 2D grasp, placement, and movement hints; a single vision-language-action model receives RGB observations and executes all task-relevant motion without carrying memory between episodes.

They augmented demonstrations with structured hint labels and trained the action model with condition dropout, spatial noise, and temporal jitter. These perturbations were intended to make execution robust when the agent's language or spatial hints are incomplete or imprecise.

Key findings

• On LIBERO-Mem, 2AM achieved 76.3% average completion, 61.5 percentage points above the strongest reported baseline of 14.8%.
• The system recorded 63.0% relaxed success and 11.8% strict success.
• These results were obtained without depth input, online geometry, or planner-driven object motion.
• The authors show that task history can remain entirely agent-side if the action model is trained to respond to sufficiently precise language and 2D spatial guidance.

Why it matters

The study isolates the role of the interface between a memory-bearing agent and a robot policy. Its results suggest that long-horizon capability need not come from adding recurrence or memory to the motor policy; it can instead emerge from a higher-bandwidth interface that communicates intent at multiple spatial and temporal scales.

Caveats

The evidence is limited to LIBERO-Mem and does not establish that the same separation will work in less controlled physical environments or across other robot embodiments. Strict success remains low at 11.8%, and the abstract does not disentangle how much each hint type or training perturbation contributes to performance.