How Agility Robotics Uses NVIDIA Isaac to Train Humanoid Robots
Agility Robotics uses NVIDIA Isaac Sim and Isaac Lab to train its Digit humanoid robot through billions of GPU-accelerated simulation steps. This simulation-first approach cut iteration cycles from weeks to days, enabling successful deployment at GXO Logistics and Schaeffler manufacturing facilities.
Impact
Weeks to days
Iteration cycle time
Billions of simulation steps
Training scale
2 enterprise sites
Production deployments
Challenge
Teaching a bipedal humanoid robot reliable whole-body control across unpredictable real-world conditions required exposing it to thousands of scenarios that were impractical to test physically.
Solution
Agility used NVIDIA Isaac Sim and Isaac Lab to simulate billions of training interactions on GPUs, reducing iteration cycles from weeks to days and enabling sim-to-real transfer at production scale.
Tools & Technologies
What Leaders Say
“Isaac Sim running on NVIDIA GPUs lets us simulate years of real-world learning for Digit in just hours. That simulation speedup means we can train for all conditions we might encounter on the factory floor.”
Sign up to read complete case studies, access detailed metrics, and unlock all use cases.
Full Story
U.S. warehouses face persistent labor shortages on material-handling lines, and remodeling facilities for fixed automation is costly and slow. Agility Robotics set out to build Digit, a general-purpose humanoid robot capable of operating in human-built spaces without facility modifications—but teaching a bipedal robot reliable whole-body control in unpredictable real-world environments proved enormously complex.
The core challenge was that real-world training for bipedal robots is prohibitively slow and expensive. Engineers needed to expose Digit to thousands of stress-test scenarios—varying aisle widths, floor friction, lighting, and unexpected shove forces—without breaking hardware or spending months on physical trials. Traditional simulation tools lacked the physics fidelity and scale required to transfer learned behaviors reliably to the real world.
Agility adopted a simulation-first approach using NVIDIA Isaac Sim, modeling Digit's chassis with OpenUSD to preserve every joint, mass, and contact surface. Customer CAD and BIM data integrated into the same scene graph, enabling realistic scenario scripting. The team then ran millions of parallel reinforcement-learning episodes in Isaac Lab, pushing training into billions of interactions across NVIDIA GPUs. To validate that controller physics was not simulator-specific, they also ran the same policy through a containerized MuJoCo pipeline, using contact-physics differences to expose corner cases and harden the model.
The results translated directly to production. Iteration cycles for developing and testing new controllers dropped from weeks to days. GXO Logistics deployed Digit fleets in a Georgia fulfillment center under a Robots-as-a-Service agreement—the world's first humanoid RaaS deployment—where robots pick totes, sort inventory, and work alongside human crews. Schaeffler, the global motion-technology supplier, deployed Digit at its Cheraw, South Carolina plant to load and unload washing-machine housings with high precision.
These deployments confirmed that skills learned in NVIDIA Isaac Sim scale from logistics totes to precision metal stampings. Agility Robotics is now developing adaptive learning systems that let each robot learn from single task demonstrations and practice autonomously in simulation, targeting deployment timescales of hours rather than weeks for new workflows.