Boston Dynamics Atlas (Electric) & Agility Digit
Duration: 40 min · Level: Foundational · Module: 1. The Humanoid Landscape · Focus: boston-dynamics, atlas, agility, digit
Figure and Tesla bet on general-purpose learning. The two companies in this lesson bet on something quieter and, in the near term, arguably more practical: task-specific deployment, done reliably, at manufacturing scale. Boston Dynamics and Agility Robotics have the deepest hardware pedigrees in the field, and in 2024 both made moves that reveal a different theory of how humanoids reach real work.
Boston Dynamics retires hydraulics
In April 2024, Boston Dynamics retired its hydraulic Atlas and introduced a fully electric successor. This is more than a component swap. The hydraulic Atlas was the most athletic robot ever built, but hydraulics are loud, leak, and demand bulky pumps — fundamentally unsuited to a commercial product working near people. Going electric trades raw theatrical power for the quiet, clean, controllable operation a workplace actually requires.
Boston Dynamics has published few hard specs for the electric Atlas, but claims it is stronger and has a greater range of motion than the hydraulic version — notably, joints that can rotate in ways a human's cannot, which can make confined-space tasks easier rather than harder. The deployment target is concrete: through its Hyundai partnership, Atlas is aimed at automotive manufacturing — spot welding, parts handling, and quality inspection on Hyundai and Kia lines.
Agility builds the first humanoid factory
Agility Robotics, also owned by Hyundai, has been less flashy and more operational. Its robot Digit — 1.75 m, 65 kg, designed for logistics — has been working in Amazon fulfillment centers handling totes since 2023. Digit's design choices are pragmatic: a bespoke MPC-based locomotion stack paired with learned manipulation policies, and legs that fold to navigate the tight aisles of a real warehouse.
The most significant thing Agility built, though, is not a robot but a building. RoboFab in Salem, Oregon, has capacity for 10,000 Digit units per year — the first humanoid-scale mass-production facility in North America. Where Tesla intends to manufacture humanoids inside its car factories, Agility built a factory whose only product is the humanoid. That is a bet that the bottleneck to adoption is not intelligence but supply.
A different theory of the problem
The defining contrast with Figure and Tesla is philosophical. Both Boston Dynamics and Agility prioritize task-specific deployment over general-purpose learning. Rather than pursuing one model that can do anything, they field robots that do a bounded set of tasks — tote handling, parts transfer — extremely reliably, and they scale by manufacturing and deploying more units against those known tasks.
Neither approach is obviously correct, and that is the point of studying both. General-purpose learning has the higher ceiling; task-specific deployment has the shorter, more certain path to revenue today. A robot that reliably moves totes for eight hours is earning money now, while general-purpose policies are still being trained.
What G1 should take from it
For G1, this lesson supplies a crucial design discipline: reliability in a bounded domain can beat brilliance in an unbounded one. Agility's folding legs and MPC locomotion are reminders that hard-won, classical engineering still matters, and that "deployable today" is itself a feature. G1's healthcare strategy — developed in Lesson 1.6 — will lean on exactly this insight: pick a domain, make the robot trustworthy and dependable within it, and let manufacturing do the rest.
Putting it into practice
Compare the two deployment philosophies on a single axis: time-to-revenue versus capability ceiling.
- Place all four Module 1 companies so far — Figure, Tesla, Boston Dynamics, Agility — on a 2×2 of "general-purpose ↔ task-specific" and "research-stage ↔ deployed-today."
- For Boston Dynamics and Agility specifically, list the engineering choices that signal a deployment-first mindset (electric over hydraulic, folding legs, MPC locomotion, a dedicated factory).
- Estimate, qualitatively, which approach reaches positive unit economics first, and which has the higher long-run ceiling. State the assumption behind each estimate.
- Decide where a safety-first healthcare humanoid like G1 belongs on your 2×2, and why that quadrant is currently uncontested.
Key takeaways
- Boston Dynamics retired hydraulic Atlas in April 2024 for an electric successor — quieter, cleaner, stronger, with greater range of motion — targeting Hyundai/Kia auto lines.
- Agility's Digit (1.75 m, 65 kg) has handled totes at Amazon since 2023 using MPC locomotion, learned manipulation, and folding legs for tight spaces.
- Agility's RoboFab (Salem, OR) can build 10,000 Digit units/year — North America's first humanoid-scale factory, betting that supply, not intelligence, is the bottleneck.
- Both firms prioritize reliable task-specific deployment over general-purpose learning: a shorter path to revenue with a lower ceiling.
- G1's lesson: reliability in a bounded domain can beat brilliance in an unbounded one — the foundation of its healthcare strategy.
← Previous: 1.3 Tesla Optimus — Scale as Moat · Next: 1.5 Unitree G1 & the Open Research Ecosystem →
Part of Module 1: The Humanoid Landscape.