AWS's Physical AI Toolchain Costs $79 to Fine-Tune GR00T, but Its Edge Deployment Stage Is Still Marked Planned

AWS published its Physical AI Toolchain on GitHub this week, and the repository's own README lists the edge-deployment stage as "Planned" while the launch coverage describes five stages. The repo is aws-samples/sample-aws-physical-ai-toolchain, in the aws-samples organisation on GitHub.

The Robot Report's Steve Crowe reported the launch on October 8. He quotes Sri Elaprolu, director of Frontier AI Science and Engineering at AWS, on the design: "The toolchain is intentionally staying neutral to that final step", meaning which robot the model ends up on. The story says the toolchain is not a replacement for RoboMaker, the simulation service AWS shut down in 2025.

What the README says is available

The README splits the work into four pillars (synthetic data, model training, simulation, sim-to-real) with NVIDIA components in each: Isaac Sim, Cosmos, Isaac Lab, GR00T N1.6 and a 14B-parameter world model called DreamZero. Orchestration runs on NVIDIA OSMO 6.3 on Amazon EKS. Eight components are marked "Available", one is "Preview" (Isaac Lab Arena evaluation) and one is "Planned".

The planned one is edge deployment: model packaging to NVIDIA Jetson through EKS Hybrid Nodes and AWS IoT Greengrass. The launch article names Greengrass as the way models reach edge devices. The README's pipeline diagram says "Deploy: export to TensorRT, deploy to robot fleet via Greengrass", yet its component table has no deployable edge module yet. A team reading the press coverage would assume the last mile ships; the repo says it does not.

The cost table is the useful part

The README carries an estimated-cost table, and it is rare for a reference architecture to publish one. The denominators matter: each figure is for a named instance and a named duration.

Component	Estimate	Instance and duration
GR00T fine-tune, full	about $79	ml.g5.12xlarge, 11 hours
DreamZero fine-tune, 1000 steps	about $93	ml.g7e.24xlarge, 4 h 11 m
Cosmos 3 Predict	about $37 an hour	p5.48xlarge, Capacity Block
Isaac Sim workstation	about $1.86 an hour	g6e.4xlarge
OSMO deployment	about $5 an hour	EKS, RDS, ElastiCache

The $79 GR00T run is a fine-tune on a UR3 arm with a Robotiq 2F-85 gripper, using 27 real teleoperation episodes. That is a demonstration dataset, not a production one, so the $79 prices the sample and not a fleet. The $37 an hour Cosmos figure assumes a reserved Capacity Block rather than on-demand capacity.

1. GR00T smoke test (15 min)2 USD
2. DreamZero smoke gate (25 min)10 USD
3. GR00T full (11 hours)79 USD
4. DreamZero 1000 steps (4 h 11 m)93 USD

Source: aws-samples/sample-aws-physical-ai-toolchain README, accessed 2026-10-10

A customer claim to check

The launch article lists Telexistence among companies using AWS, describing it as deploying humanoids in convenience stores, with more than 300 reported. Robotics 24/7's 2022 coverage of Telexistence describes something different: the TX SCARA, a track-mounted arm that is explicitly not humanoid, starting in 300 FamilyMart stores from August 11, 2022, with remote operators able to take over through a VR system.

The two may describe different products; Telexistence could have a newer machine. We found no source that says so. Until one appears, the 300 figure is a 2022 store count for a rail-mounted arm, not a fleet of humanoids.

What it means for buyers

The toolchain is hardware-neutral by design, which suits a market where Jabil said humanoids are entering tens of thousands without a customer or a count, and where the FCC's foreign robot rule may decide which components a buyer can use. An $79 fine-tune is small next to a robot; the expense sits in the teleoperation data and the machine itself.

The number to watch is the repository's status table: when the edge-deployment row moves from "Planned" to "Available", the pipeline becomes closed-loop in fact as well as in the diagram. Until then, the cheapest part of the stack is the part AWS has shipped.