Humanoid robotics pilot

Humanoids are only useful if they can actually do work. This pilot pushed past the demo-gif and asked what whole-body locomanipulation looks like inside an enterprise operating environment.

Problem

A Japanese industrial conglomerate wanted to evaluate humanoid robotics for real operational use — not a marketing film. That required a credible path from policy training in simulation to behavior that transfers to a physical humanoid in their environment, plus a general-purpose orchestration layer that non-robotics operators could actually reason about.

Approach

Led simulation design and owned the solution end-to-end. Trained whole-body locomanipulation behaviors in NVIDIA Isaac Lab — the learning environment of choice for modern humanoid policies — and used Google Gemini Robotics as the general-purpose orchestration layer on top, so high-level task intent could be mapped to low-level learned skills. Designed the program around sim-to-real transfer, enterprise integration, and the question that actually matters to clients: where in the workflow does a humanoid add more value than a fixed arm or an AMR.

Outcome

• Live pilot with Japanese conglomerate using Isaac Lab policies + Gemini Robotics orchestration
• Reference architecture for enterprise humanoid evaluation, reusable across Slalom’s humanoid engagements
• TODO: verify metric — number of trained skills / behaviors
• TODO: verify metric — sim-to-real success rate
• TODO: verify metric — pilot duration and next-phase commitment