Software Development Engineer II, Amazon Robotics - Manipulation

About the role

The Fleet Performance Optimization (FPO) organization builds the intelligence layer for Amazon Robotics manipulation workcells. We own the systems that select what work a robot should attempt, monitor how it performs, and close the loop by improving models with every induct. Our platforms serve Sparrow, Cardinal, FlexCell, and Robin workcells that handle millions of packages across Amazon's fulfillment network.

We operate across the full ML and data lifecycle:

Work Selection Intelligence — Eligibility scoring and predictions (likelihood-to-empty, damage prediction, time-to-process, pick eligibility) that feed into work planners, identifying which totes and items reach each workcell.

Data & Observability — A common datalake collecting telemetry from every automated and manual workcell, fleet monitoring that detects performance degradation in real time and alerts Ops, and deep-dive tooling that enables engineers and scientists to investigate workcell behavior.

ML Lifecycle — Annotation orchestration, model training pipelines, production monitoring, and deployment infrastructure that keeps models current across the fleet.

Predictive Models — Science-developed foundational models applied to eligibility, fleet performance forecasting, and dynamic floor policies that adapt to changing conditions without manual intervention.

Our systems process tens of millions of events daily, serve real-time scoring at workcell-runtime latency, and directly impact fulfillment cost per unit.

Key job responsibilities

As an SDE on FPO, you will design, build, and operate distributed systems that sit at the intersection of robotics, machine learning, and large-scale data processing. You will work closely with scientists, program facilators, and partner engineering teams to translate research into production systems that operate reliably at fleet scale.

Depending on the team, your work may include:

- Building high-throughput, event-based scoring services that process real-time inventory signals across 10+ warehouses and hundreds of workcells

- Designing config-based annotation orchestration systems that onboard new ML data pipelines without code changes

- Developing model deployment infrastructure spanning edge devices, cloud inference services, and planning systems

- Building fleet monitoring systems that use VLMs and statistical methods to detect performance anomalies and surface root causes

- Creating unified data exploration and visualization tools used daily by engineers, scientists, and operations

- Implementing ongoing learning systems that intelligently select which data to collect based on inference outcomes

- Extending our foundation models, which are the backbone for multiple prediction tasks

- Own your systems end-to-end — from design through production operations — and your work will have direct, measurable impact on workcell throughput, quality, and cost.