The core change is breadth of control and task horizon. A single VLA can walk, crouch, balance, and manipulate—reducing brittle handoffs between locomotion and manipulation stacks. The ER layer plans multi-minute sequences, tracks events that mark progress, and coordinates different robots to divide work. Practically, this lets you scope pilots around full workflows—room tidying, shelf restocking, or kit packing—rather than building glue logic across separate controllers. It also compresses integration time: fewer APIs, fewer failure points, and one telemetry surface for reasoning and control.

- Scope one constrained workflow with clear guardrails: e.g., pick, pack, and stage totes on two shelf levels. - Hardware: one humanoid or bi-arm base plus a gripper and a five-finger hand to compare dexterity trade-offs. - Data: collect 150–250 high-quality episodes mixing teleop and autonomous rollouts; log interventions with reasons. - Evaluation: define target cycle time, first-pass success (no regrasp), recovery latency, near-miss rate, and safe stop frequency. - Safety: require prox sensors, geofencing, and agentic refusal checks, with operator takeover under two seconds.

Multifinger dexterity remains inconsistent for fine assembly and deformables, and longer sequences can drift without reliable event detection. On-device models mitigate latency but raise scheduling and thermal constraints; expect occasional policy slowdowns under high load. Safety orchestration must refuse unsafe motions, predict task impossibility, and escalate to humans early. Adopt red-teaming for proximity, pinch, and topple scenarios; require immutable safety layers (E-stops, torque limits) independent of the learning stack; and log safety tool calls for audit. Regulatory alignment for collaborative operation will demand documented hazard analyses and repeatable test suites.