Agility Robotics Digit — Architecture Analysis

Commercial-Live

Agility Robotics shipped the first commercial humanoid deployment in history at GXO Logistics in June 2024 ^[1]. A 2015 spin-out from the Oregon State University Dynamic Robotics Lab, Agility brings a decade of biomimetic locomotion research (ATRIAS, Cassie) into a warehouse-focused humanoid product. Where Tesla bets on vertical integration and Figure on an in-house foundation model, Agility bets on the deepest non-Boston-Dynamics locomotion R&D lineage and a commercial-first business model (Robots-as-a-Service).

Body DOF

19

Bipedal-only

Payload

16 kg

Per arm

Height

175 cm

5'9"

Weight

65 kg

Heaviest of cohort

Runtime

4+ hr

Hot-swap pack

Speed

1.5 m/s

Walking, not run

Commercial milestone: June 2024 GXO Logistics deployment is the first humanoid in history to do real, billed, multi-shift warehouse work for a paying customer ^[1]. Tesla, Figure, Apptronik, and Boston Dynamics all remain in pilot or internal-deployment stages as of writing. Agility's lead here is measured in production hours, not just press releases.

No public teardown: Agility publishes deployment milestones and high-level specs but does NOT publish actuator model numbers, peak torque tables, battery cell vendors, or contract-manufacturing partners. Cassie heritage is well-documented in OSU academic papers; the productised Digit stack is far less public. Spec gaps below are flagged “Not publicly disclosed” rather than guessed.

Backwards-knee design:Digit's most visible mechanical signature is its reverse-knee bipedal leg — biomimetic ostrich/emu kinematics inherited from OSU ATRIAS (2013) and Cassie (2017) ^[2]. The knee bends backwards relative to human anatomy, optimising walking energy economy. This is a deliberate departure from human-anatomical bipeds (Optimus, Figure, Atlas, Apollo). It divides opinion — some argue it's not a humanoid; Agility argues function beats form.

Approach

First commercial humanoid deployment in history (GXO Logistics, June 2024). RaaS business model. Biomimetic lower-body design.

•FIRST commercial humanoid deployment in history (GXO/Spanx, June 2024)

•RoboFab: 70,000 sq ft manufacturing, 10,000 units/year capacity, Salem OR

•Amazon pilot at Seattle R&D — tote recycling tasks

•Oregon State University biomechanics heritage

•DCVC + Playground Global backing + Amazon Industrial Innovation Fund

Component-by-Component Teardown

Each Digit subsystem mapped to the closest commodity equivalent from our verified parts catalog. Where Agility has not disclosed a spec, the row is labelled Not Disclosed.

Subsystem	Digit Implementation	Public Spec	Commodity Equivalent	Price	Gap
Hip / Shoulder Actuator	Agility-designed series-elastic actuator (SEA), derived from OSU Cassie heritage; specific model numbers not disclosed	Cassie-class SEA: brushless DC + planetary reducer + leaf-spring elastic element + torque sensor. Per-joint torque/RPM tables not published.	Harmonic FHA-25C (108 Nm) or T-Motor RMD-X8 (80 Nm)	$660–$2,400	Not Disclosed
Knee — Backwards-Knee Geometry	Reverse-knee bipedal leg, biomimetic ostrich/emu-inspired kinematics (Cassie heritage)	Knee bends backwards relative to human anatomy — heel-equivalent is the ankle joint, ankle-equivalent is the toe. Optimised for dynamic walking efficiency.	T-Motor RMD-X8 (80 Nm) — geometry-agnostic rotary	$620–$1,650	Different Approach
End Effector / Hand	End-effector hardware has evolved: early Digit used 4-finger paddle grippers; current production units have transitioned to articulated hands. Specific DOF count and design generation not consistently published.	Payload 16 kg per arm (published). Hand DOF: not publicly fixed at the time of writing. Tactile sensing: not publicly disclosed.	Dynamixel XL330 array — servo-per-DOF approximation	$65/servo	Not Disclosed
Compute	On-board compute — NVIDIA-class (graph.ts pairs agility with nvidia-thor). Specific module not publicly named.	Believed Jetson AGX Thor or Orin-class. Cassie's research stack used Intel NUC + ROS — Digit's production compute is presumed upgraded but not detailed in public materials.	NVIDIA Jetson AGX Thor (2,070 FP4 TFLOPS, 130W)	$3,499	Close Match
Battery	Third-party Li-ion pack — vendor not publicly disclosed	Runtime 4+ hours (companies.ts spec). Pack architecture, cell chemistry, BMS designer not in the public record.	Victron Lithium 48V 50Ah (2.4 kWh)	$3,200	Not Disclosed
Vision — Stereo + LiDAR	Stereo cameras + LiDAR (Intel RealSense reported in earlier prototypes; LiDAR class not specified)	Perception stack designed around perceptive step planning — a Cassie heritage. Sees terrain, plans footfall, balances dynamically.	Intel RealSense D455 + LiDAR (Slamtec RPLIDAR S2)	$400–$800 + ~$650 LiDAR	Partial Gap
Force/Torque Sensing	Force/torque sensors at wrists and feet (graph.ts: agility → force-torque)	6-axis F/T at end-effectors and contact points. Critical for tote handling and dynamic walking foot-placement.	ATI Mini-40 6-axis F/T sensor	$2,400	Close Match
Frame / Structure	Aluminum + composite structural frame. Manufacturing at RoboFab (Salem, OR), 70,000 sq ft facility, stated 10,000 units/year capacity.	65 kg total mass, 175 cm tall. No Giga-Press analog. Final assembly in-house at Salem.	Carbon fiber mid-tier or aluminum frame	$3,500–$22,000	Partial Gap

Hip / Shoulder Actuator

Not Disclosed

Digit

Agility-designed series-elastic actuator (SEA), derived from OSU Cassie heritage; specific model numbers not disclosed

Cassie-class SEA: brushless DC + planetary reducer + leaf-spring elastic element + torque sensor. Per-joint torque/RPM tables not published.

Commodity

Harmonic FHA-25C (108 Nm) or T-Motor RMD-X8 (80 Nm)

$660–$2,400

Digit's actuator family descends directly from the OSU Cassie research robot (2017) — series-elastic, brushless DC, planetary-stage reduction. Compliance is intrinsic to the design (the elastic element absorbs impact and stores energy during walking). Agility has NOT published model numbers, peak torque per joint, or supplier names. Closest commodity proxies: harmonic-drive rotaries for stiffness, T-Motor / CubeMars quasi-direct-drive for compliance. Neither is a true SEA — the spring element is what makes Cassie/Digit walking efficient.

Knee — Backwards-Knee Geometry

Different Approach

Digit

Reverse-knee bipedal leg, biomimetic ostrich/emu-inspired kinematics (Cassie heritage)

Knee bends backwards relative to human anatomy — heel-equivalent is the ankle joint, ankle-equivalent is the toe. Optimised for dynamic walking efficiency.

Commodity

T-Motor RMD-X8 (80 Nm) — geometry-agnostic rotary

$620–$1,650

The backwards-knee design is Digit's most visible architectural signature — and a deliberate departure from human-anatomical bipeds (Optimus, Figure, Atlas, Apollo). Roots: Jonathan Hurst's research at Oregon State Dynamic Robotics Lab, where ATRIAS (2013) and Cassie (2017) used point-feet ostrich-style legs to maximise walking energy economy. Digit retains the leg topology but adds a torso, arms, and end-effectors. Commodity rotary actuators can replicate the joint torques; the kinematic chain is the differentiator and is well-documented in OSU papers.

End Effector / Hand

Not Disclosed

Digit

End-effector hardware has evolved: early Digit used 4-finger paddle grippers; current production units have transitioned to articulated hands. Specific DOF count and design generation not consistently published.

Payload 16 kg per arm (published). Hand DOF: not publicly fixed at the time of writing. Tactile sensing: not publicly disclosed.

Commodity

Dynamixel XL330 array — servo-per-DOF approximation

$65/servo

Agility's end-effector strategy is warehouse-driven: grasp totes, lift packages, place items on conveyors. This is a far simpler manipulation envelope than Optimus's 22-DOF tendon hands or Figure's 16-DOF tendon hands. Digit has historically prioritised payload (~16 kg) over manipulation dexterity. The exact current production hand spec (DOF count, tactile sensing, gripper-vs-hand) has not been consistently published and varies between generations.

Compute

Close Match

Digit

On-board compute — NVIDIA-class (graph.ts pairs agility with nvidia-thor). Specific module not publicly named.

Believed Jetson AGX Thor or Orin-class. Cassie's research stack used Intel NUC + ROS — Digit's production compute is presumed upgraded but not detailed in public materials.

Commodity

NVIDIA Jetson AGX Thor (2,070 FP4 TFLOPS, 130W)

$3,499

Knowledge-graph wiring (graph.ts: agility → nvidia-thor) places Digit on the same NVIDIA-class compute stack as Figure, Apptronik, UBTech. NVIDIA is a confirmed industry partner across the warehouse-humanoid cohort. Agility has not confirmed the SKU. Locomotion stack (the harder problem) is well-understood from OSU Cassie publications; perception + foundation-model strategy is less public.

Battery

Not Disclosed

Digit

Third-party Li-ion pack — vendor not publicly disclosed

Runtime 4+ hours (companies.ts spec). Pack architecture, cell chemistry, BMS designer not in the public record.

Commodity

Victron Lithium 48V 50Ah (2.4 kWh)

$3,200

Agility states 4+ hour runtime per charge with hot-swappable battery packs (key for warehouse multi-shift operation). Cell vendor not announced — likely candidates: LG Energy, Samsung SDI, CATL, BYD, or a US-domiciled supplier given Albany OR / Salem OR manufacturing footprint. Hot-swap pack architecture is publicly demonstrated in GXO deployment videos.

Vision — Stereo + LiDAR

Partial Gap

Digit

Stereo cameras + LiDAR (Intel RealSense reported in earlier prototypes; LiDAR class not specified)

Perception stack designed around perceptive step planning — a Cassie heritage. Sees terrain, plans footfall, balances dynamically.

Commodity

Intel RealSense D455 + LiDAR (Slamtec RPLIDAR S2)

$400–$800 + ~$650 LiDAR

Unlike Tesla's vision-only doctrine and Figure's RGB-first VLA approach, Digit retains LiDAR — consistent with its locomotion-first heritage. OSU Cassie used Intel RealSense for terrain perception in published research. Production Digit configuration not exhaustively documented but stereo + LiDAR is consistent across press materials and demo videos.

Force/Torque Sensing

Close Match

Digit

Force/torque sensors at wrists and feet (graph.ts: agility → force-torque)

6-axis F/T at end-effectors and contact points. Critical for tote handling and dynamic walking foot-placement.

Commodity

ATI Mini-40 6-axis F/T sensor

$2,400

Force/torque sensing is explicit in graph.ts wiring (agility → force-torque) and a known requirement for warehouse manipulation tasks. ATI Industrial is the dominant Western F/T sensor supplier — likely though not confirmed source. Cassie's research stack used custom load cells; Digit's production sensors are presumably commercial.

Frame / Structure

Partial Gap

Digit

Aluminum + composite structural frame. Manufacturing at RoboFab (Salem, OR), 70,000 sq ft facility, stated 10,000 units/year capacity.

65 kg total mass, 175 cm tall. No Giga-Press analog. Final assembly in-house at Salem.

Commodity

Carbon fiber mid-tier or aluminum frame

$3,500–$22,000

RoboFab opened September 2023 as the first dedicated humanoid robot factory in the US. Stated 10K/yr capacity. Frame is machined and assembled rather than die-cast at Tesla's Giga-Press scale. Materials decomposition not publicly broken out. Mass budget (65 kg at 175 cm) is heavier than Figure (60 kg at 165 cm) and Optimus (57 kg at 173 cm) — reflecting payload-optimised design.

Gap Analysis Detail

Hip / Shoulder ActuatorNot Disclosed

Digit's actuator family descends directly from the OSU Cassie research robot (2017) — series-elastic, brushless DC, planetary-stage reduction. Compliance is intrinsic to the design (the elastic element absorbs impact and stores energy during walking). Agility has NOT published model numbers, peak torque per joint, or supplier names. Closest commodity proxies: harmonic-drive rotaries for stiffness, T-Motor / CubeMars quasi-direct-drive for compliance. Neither is a true SEA — the spring element is what makes Cassie/Digit walking efficient.

Knee — Backwards-Knee GeometryDifferent Approach

The backwards-knee design is Digit's most visible architectural signature — and a deliberate departure from human-anatomical bipeds (Optimus, Figure, Atlas, Apollo). Roots: Jonathan Hurst's research at Oregon State Dynamic Robotics Lab, where ATRIAS (2013) and Cassie (2017) used point-feet ostrich-style legs to maximise walking energy economy. Digit retains the leg topology but adds a torso, arms, and end-effectors. Commodity rotary actuators can replicate the joint torques; the kinematic chain is the differentiator and is well-documented in OSU papers.

End Effector / HandNot Disclosed

Agility's end-effector strategy is warehouse-driven: grasp totes, lift packages, place items on conveyors. This is a far simpler manipulation envelope than Optimus's 22-DOF tendon hands or Figure's 16-DOF tendon hands. Digit has historically prioritised payload (~16 kg) over manipulation dexterity. The exact current production hand spec (DOF count, tactile sensing, gripper-vs-hand) has not been consistently published and varies between generations.

ComputeClose Match

Knowledge-graph wiring (graph.ts: agility → nvidia-thor) places Digit on the same NVIDIA-class compute stack as Figure, Apptronik, UBTech. NVIDIA is a confirmed industry partner across the warehouse-humanoid cohort. Agility has not confirmed the SKU. Locomotion stack (the harder problem) is well-understood from OSU Cassie publications; perception + foundation-model strategy is less public.

BatteryNot Disclosed

Agility states 4+ hour runtime per charge with hot-swappable battery packs (key for warehouse multi-shift operation). Cell vendor not announced — likely candidates: LG Energy, Samsung SDI, CATL, BYD, or a US-domiciled supplier given Albany OR / Salem OR manufacturing footprint. Hot-swap pack architecture is publicly demonstrated in GXO deployment videos.

Vision — Stereo + LiDARPartial Gap

Unlike Tesla's vision-only doctrine and Figure's RGB-first VLA approach, Digit retains LiDAR — consistent with its locomotion-first heritage. OSU Cassie used Intel RealSense for terrain perception in published research. Production Digit configuration not exhaustively documented but stereo + LiDAR is consistent across press materials and demo videos.

Force/Torque SensingClose Match

Force/torque sensing is explicit in graph.ts wiring (agility → force-torque) and a known requirement for warehouse manipulation tasks. ATI Industrial is the dominant Western F/T sensor supplier — likely though not confirmed source. Cassie's research stack used custom load cells; Digit's production sensors are presumably commercial.

Frame / StructurePartial Gap

RoboFab opened September 2023 as the first dedicated humanoid robot factory in the US. Stated 10K/yr capacity. Frame is machined and assembled rather than die-cast at Tesla's Giga-Press scale. Materials decomposition not publicly broken out. Mass budget (65 kg at 175 cm) is heavier than Figure (60 kg at 165 cm) and Optimus (57 kg at 173 cm) — reflecting payload-optimised design.

What Would It Cost to Build a Digit-Equivalent?

Running the closest commodity BOM through the HUMA Build Compiler — targeting 175 cm, 65 kg, 16 kg payload, 4 hr runtime, pro compute, 19 body DOF (Digit publishes a low-DOF spec compared to Optimus 28 or Atlas 56).

DOES NOT COMPILE

6 errors5 warnings

Total Cost

$41,695

Digit Sticker Price

~$175,000

Total Mass

36.9 kg

Peak Power

18,555.35 W

A commodity Digit-equivalent costs $41,695 vs Digit's ~$175,000 sticker price (companies.ts:agility-digit.spec.price). Note: Agility primarily sells access to Digit via RaaS rather than units — list price is a reference, not the primary go-to-market. Commodity hardware now beats Digit on price alone. Agility's defensible asset is the locomotion stack and the GXO production-hours moat, not the BOM.

✗

Electrical

Main-bus voltage mismatch: battery is 48V but 1 main-bus actuators expect different voltage

Dynamixel XM430-W350-T expects 12V

✓

Electrical

Power budget OK: 3507W effective peak vs 7200W battery

Actuator peak 11212W × 30% + 144W non-actuator

!

Communication

Multiple actuator protocols: EtherCAT, CAN, UART — requires a multi-protocol gateway

✓

Communication

Compute (NVIDIA Jetson AGX Thor (T5000 Module)) bridges to actuator bus via motor controller

!

Communication

EtherCAT actuators detected with standard Linux compute. EtherCAT requires a real-time kernel (PREEMPT_RT patch). Stock Jetson Ubuntu will have jitter issues without RT patches. Consider: (a) apply PREEMPT_RT to Jetson, (b) use a dedicated EtherCAT master (Beckhoff), or (c) switch actuators to CAN protocol.

✗

Mechanical

Missing hip or knee actuators

✓

Mass Budget

Mass budget OK: 38.9 kg / 65 kg (60%)

✗

Software

Compute too weak for pro tier: 0 TOPS < 200 TOPS required

Full VLA models (Helix, GR00T N1.5, π0.5)

!

Thermal

Thermal budget tight: 784W simultaneous heat vs 30W passive (26× — active cooling required)

Harmonic Drive FHA-25C ×4: 324W each (η=73%); T-Motor RMD-X8 Gimbal Motor ×2: 302W each (η=82%); CubeMars AK10-9 QDD Actuator ×2: 230W each (η=80%); EYOU PH11-101 Harmonic Actuator ×2: 72W each (η=70%)

!

Thermal

Low-efficiency actuator: Dynamixel XH540-W270-T at 60% — monitor joint temperature in continuous use

!

Runtime

Runtime marginal: 3.4h estimated vs 4h target

Battery 2400 Wh ÷ avg 704W (actuators 561W at 5% duty + 144W always-on). Power management or reduced duty may be needed.

✓

DOF

DOF OK: 21 actuators for 19 target DOF

✗

infrastructure

No DC-DC converter or PSU producing the 5 V rail (load 16 W)

Loads on this rail: Dynamixel XL330-M288-T ×8, Intel RealSense D455 Depth Camera ×2, Slamtec RPLIDAR S2 2D LiDAR ×1, Xsens MTi-3 AHRS ×1, ATI Mini40 F/T Sensor ×4. Add a 48→5 V converter (e.g. Pololu D36V50F-series, Mean Well SD-100C, Murata UWE/OKL).

✗

infrastructure

No DC-DC converter or PSU producing the 12 V rail (load 131 W)

Loads on this rail: Dynamixel XM430-W350-T ×1, NVIDIA Jetson AGX Thor (T5000 Module) ×1. Add a 48→12 V converter (e.g. Pololu D36V50F-series, Mean Well SD-100C, Murata UWE/OKL).

✗

infrastructure

No DC-DC converter or PSU producing the 24 V rail (load 10 W)

Loads on this rail: Dynamixel XH540-W270-T ×2. Add a 48→24 V converter (e.g. Pololu D36V50F-series, Mean Well SD-100C, Murata UWE/OKL).

i

Availability

All 14 parts have supplier URLs — stock & lead-time NOT verified

In-House vs Partner-Led

Agility's vertical is locomotion + manufacturing: the company owns the Cassie-heritage actuator design + walking controllers and assembles units in-house at RoboFab Salem. It partners on compute (NVIDIA-class), batteries (third-party Li-ion), and customer integration (GXO, Amazon).

Agility Owns / Makes In-House

Series-elastic actuator design (Cassie heritage)

Agility's actuator IP descends from a decade of Oregon State Dynamic Robotics Lab research. Founders Jonathan Hurst, Damion Shelton, Mikhail Jones brought the SEA + biomimetic-leg approach into the company at 2015 spin-out. Manufacturing is in-house at RoboFab Salem; specific motor + reducer suppliers not disclosed.

Backwards-knee bipedal kinematics

The defining mechanical signature. Roots in OSU ATRIAS (2013) and Cassie (2017) point-feet ostrich-style legs. Digit added torso + arms while preserving the energy-efficient leg topology. Documented in published OSU papers; geometry is public, control software is proprietary.

Walking + balance control stack (Cassie research lineage)

Agility's locomotion controllers descend from OSU Cassie research that demonstrated 5K runs, stair climbing, and field-deployable walking. This is one of the deepest locomotion R&D lineages outside Boston Dynamics.

RoboFab manufacturing facility (Salem, OR)

70,000 sq ft, opened September 2023, stated 10,000 units/year capacity. First dedicated humanoid robot factory in the US. Final assembly is Agility-internal — no contract manufacturer disclosed.

Robots-as-a-Service (RaaS) commercial model

Agility pioneered the RaaS pricing structure for humanoids — customers pay per hour of robot operation rather than capex purchase. Active at GXO Logistics. Sets the pricing template the rest of the industry now references.

Partners / Confirmed Relationships

GXO Logistics — first commercial customer

Agility announced the first commercial humanoid deployment in history at GXO Logistics in June 2024 (per companies.ts:agility-digit.keyPartners and Agility press release dated June 2024). Site: Spanx warehouse, GXO-operated. Tasks: tote handling, package movement. RaaS billing — Agility owns the robots, GXO pays per shift. Multi-shift continuous operation is the publicly-stated milestone. This is the single most important deployment in the commercial humanoid market to date.

Amazon Industrial Innovation Fund — strategic investor + pilot customer

Amazon IIF participated in Agility's 2024 Series D extension (companies.ts:agility-digit.funding[4]). Digit is in pilot at Amazon's Seattle R&D facility on tote recycling and material handling. Amazon is both an investor (~$150M+ extension round) and a deployment partner — the same playbook Amazon has run with Covariant, Rivian, and other strategic robotics bets.

DCVC + Playground Global — long-running financial leads

DCVC has led Agility's Series A (2018), Series B (2020), and Series C (2022 — $150M). Playground Global co-led Series B. This is one of the most consistent VC syndicates in humanoid robotics. DCVC's deep-tech thesis predates the post-ChatGPT humanoid wave.

Oregon State University — research lineage + talent pipeline

Agility was a 2015 spin-out from the OSU Dynamic Robotics Lab. Founders Jonathan Hurst, Damion Shelton, Mikhail Jones; OSU continues to produce graduate-level locomotion research that feeds Agility hiring. Cassie remains the open research platform; Digit is the productised descendant.

NVIDIA — compute partner

graph.ts:agility → nvidia-thor links Digit to the industry-standard NVIDIA Thor / Jetson compute stack. Specific module not publicly confirmed by Agility but consistent with industry-wide convergence on Jetson-class on-board compute.

Apptronik — adjacent peer (graph.ts adjacency)

graph.ts wires agility → apptronik as adjacent USA-based humanoid players. Both target warehouse / logistics and use SEA-class actuators. Not a confirmed commercial partnership but a structural peer relationship in the US humanoid landscape.

Neumann's Replicability Lens

~75%

P% (Producibility)

RoboFab Salem assembles at stated 10K/yr capacity; component manufacture is outsourced

~65%

M% (Manufacturability)

Conventional assembly — no Giga-Press analog. Volume manufacturing is stated but not yet demonstrated.

~25%

C% (Commodity Replicability)

Geometry is public (OSU papers); SEA + Cassie controllers are the defensible asset, plus GXO production-hours moat.

Funding & Captable

Agility has raised $500M+ cumulative across five rounds, reaching an estimated ~$1.4B valuation (companies.ts:agility-digit). The captable pairs deep-tech VCs (DCVC consistent across A/B/C, Playground Global on B) with a single anchor strategic at extension stage (Amazon Industrial Innovation Fund) ^[3].

2015Spin-out from Oregon State

Seed

Lead: —

2018Series A

$8M

Lead: DCVC

2020Series B

$24M

Lead: DCVC, Playground Global

2022Series C

$150M

Lead: DCVC

2024Extension/D

$150M+

Lead: Amazon Industrial Innovation Fund

Why ~$1.4B matters: Compared to Figure's $39B, Agility trades at roughly 3.5% of Figure's paper-equity— yet it has already shipped the first commercial humanoid deployment in history. The market is pricing Figure on a foundation-model bet and Agility on a steady commercial-revenue trajectory. If Helix-class VLA capability becomes generic, Agility's GXO moat (production hours, customer trust, RaaS billing infrastructure) compounds — and the valuation gap narrows.

Customer Story — GXO Logistics (Spanx, June 2024)

The single most important deployment in commercial humanoid history. Agility and GXO Logistics announced multi-year RaaS agreement in June 2024, deploying Digit at a Spanx-fulfilment warehouse operated by GXO ^[1].

Deployment Profile

Site: GXO-operated Spanx fulfilment warehouse, USA
Tasks: Tote handling, package movement, conveyor-line transfer
Shift profile: Multi-shift continuous operation (~7 hour shifts publicly cited; hot-swap battery between shifts)
Billing: RaaS — GXO pays per hour of robot operation, Agility owns the hardware
Announced: June 2024 (Agility + GXO joint press release)

Why It's a Milestone

First, period: The first humanoid in history doing real billed warehouse work for a paying enterprise customer
Validated RaaS: Agility's pricing model has now been proven at a tier-1 3PL — the rest of the cohort is pricing-following
Behavioural envelope: Demonstrates that 19 DOF + 16 kg payload is enoughfor a real economic task. You don't need 28 DOF or 22-DOF hands to extract real warehouse value.
Defensible moat: Every hour Digit operates at GXO is data + customer-trust no competitor has. Tesla, Figure, Apptronik are still in pilot.

The implication for kit builders:Digit's GXO win proves the warehouse market doesn't need a 22-DOF tendon-driven hand to monetise. A 19-DOF biped with a simple gripper and 16 kg payload is enough. That is a much lower hardware bar than building an Optimus or Figure clone — and the manipulation gap can be closed incrementally with software.

Digit vs Optimus vs Figure — Side by Side

Three philosophies from the same year. Tesla owns the hardware floor; Figure owns the AI ceiling; Agility owns the customer-base.

Dimension	Tesla Optimus	Figure 02 / 03	Agility Digit
Strategy	Vertical integration	AI-first, hardware partnered	Locomotion-first, commercial deployment
Body DOF	28	24	19
Hand strategy	22 DOF tendon (Gen 3)	16 DOF tendon	Low-DOF gripper / hand (evolving)
Leg geometry	Human-anatomical	Human-anatomical	Reverse-knee, biomimetic ostrich
Payload	20 kg target	20 kg target	16 kg per arm (delivered)
Compute	Custom AI4 / AI5 (Tesla)	NVIDIA Jetson-class	NVIDIA Jetson-class
Foundation model	End-to-end NN, FSD-derived	Helix VLA (in-house)	Locomotion controllers + perception (Cassie heritage); foundation-model strategy not public
First customer	Internal Tesla factories	BMW Spartanburg pilot	GXO Logistics — first commercial humanoid in history
Pricing model	Direct sale, <$20K target	Direct sale, $50K target	RaaS — pay per hour
Manufacturing	Giga Press, Fremont	BotQ Sunnyvale 12K/yr	RoboFab Salem 10K/yr (opened Sep 2023)
Valuation	Part of $800B+ Tesla	$39B private	~$1.4B private

Where Digit is ahead

•First commercial humanoid deployment in history (GXO, Jun 2024)
•Deepest non-Boston-Dynamics locomotion R&D lineage (OSU, decade-plus)
•Validated RaaS pricing model — the rest of the cohort is following
•Production hours moat at GXO — months of real billed work
•Backwards-knee design: walking energy efficiency advantage
•Lower DOF target = lower BOM cost at equivalent payload

Where Optimus / Figure are ahead

•Higher hand DOF (22 / 16 vs Digit's low-DOF gripper)
•More mature foundation-model story (Helix VLA, end-to-end FSD-derived NN)
•Higher unit-economics ceiling at scale (target $20K / $50K vs Digit ~$175K sticker)
•Larger raise + valuation = longer runway for hardware iteration
•Human-anatomical form factor reads more naturally to general-public market
•Tesla's Giga-Press manufacturing scale advantage at 100K+ unit volume

Open Questions for the Portal

Items where Agility has not made the public record. Surface these here so visitors know what we know we don't know — and so the community can fill in gaps as teardowns and academic publications appear.

Q1.

What are the actuator model numbers and peak torque tables?

Agility's SEAs descend from OSU Cassie (well-documented in academic papers — Hurst, Fern, Hubicki et al.) but the productised Digit actuators have NOT had model numbers, peak torque per joint, or supplier names disclosed. This is the largest single spec gap. Cassie publications give a research-grade reference; production specs likely diverge.

Q2.

Who supplies the battery cells and pack architecture?

Cell vendor, chemistry, BMS designer not on record. Hot-swap pack architecture is publicly demonstrated. 4+ hour runtime at humanoid power profile (~400-500 W mixed) implies ~2-3 kWh per pack. Likely candidates: LG Energy, Samsung SDI, CATL, BYD, or a US-domiciled supplier given Salem OR manufacturing footprint.

Q3.

Who is the contract manufacturer for actuator components?

RoboFab does final assembly. Upstream — motor windings, harmonic / planetary reducers, encoder packs — Agility has not announced a CM partnership. Closest analogues: Tesla uses Sanhua + Tuopu + Green Harmonic; Figure uses undisclosed; Agility likely uses a mix of US + Asian suppliers given the US-flag investor base.

Q4.

What is Agility's foundation-model strategy?

Locomotion + balance is well-understood from Cassie heritage. Manipulation, language understanding, and end-to-end task learning are far less public. Agility has not announced an in-house VLA model (Helix-class) nor a foundation-model partnership (Physical Intelligence, NVIDIA GR00T, Google Gemini Robotics). This is the biggest strategic blank.

Q5.

What is the exact Digit V5 / V6 hand specification?

End-effector hardware has evolved across generations. Early Digit used 4-finger paddle grippers; current production has transitioned to articulated hands. Specific DOF count, tactile sensing modality, and design generation (V5 vs V6) are not consistently published. Material for a future teardown.

Q6.

What are the per-shift unit economics at GXO?

RaaS billing rate per hour, GXO contract length, and Agility's gross margin per Digit-shift have not been disclosed. These numbers will define whether RaaS is a sustainable model — and whether Tesla's direct-sale-at-$20K can undercut it on a per-task-cost basis.

Q7.

Compute SKU on-board?

graph.ts wires agility → nvidia-thor but Agility has not confirmed Thor specifically. Could be Jetson AGX Orin (current generation), Thor (next generation), or a multi-Orin configuration. Has implications for thermal envelope and on-board model size.

Q8.

How does Digit handle the Cassie point-foot to humanoid-foot transition?

Cassie used point-feet (no toe articulation). Digit added a foot for stability + payload but kept the backwards-knee leg topology. The exact ankle / foot mechanism has not been deeply documented in production materials. OSU papers cover the research lineage; Digit's production foot may have evolved further.

Why This Matters for Kit Builders

Lower DOF can be enough

Digit ships with 19 body DOF — 9 fewer than Optimus, 5 fewer than Figure — and is doing real billed work at GXO. The warehouse market does not require a 22-DOF tendon hand. Kit builders targeting logistics can ship with simpler grippers and still compete on the economic-task envelope.

Locomotion is the floor, not the ceiling

Cassie heritage gives Digit a decade-plus lead on walking efficiency. Open-source equivalents exist (MuJoCo Cassie models, Berkeley humanoid stack) but the production-grade tuning and field-deployment maturity are non-trivial. Walking well is harder than it looks; Cassie publications are the public on-ramp.

RaaS is the proven path to revenue

Selling robots is hard. Selling robot-hours is easier — the customer pays operating cost, you keep the asset. Agility proved this at GXO. Kit builders who sell as RaaS rather than $50K hardware boxes have a structurally easier go-to-market.

Sources & Footnotes

Agility Robotics + GXO Logistics joint press release, June 2024 — first commercial multi-year RaaS humanoid deployment in history. Site: GXO-operated Spanx fulfilment warehouse, USA. Captured in companies.ts:agility-digit.keyPartners and companies.ts:agility-digit.strengths.
Oregon State University Dynamic Robotics Lab — ATRIAS (2013) and Cassie (2017) academic publications (Hurst, Fern, Hubicki et al.). Open-access papers describe the biomimetic backwards-knee leg geometry and series-elastic actuator architecture that Digit productises.
Funding rounds per companies.ts:agility-digit.funding — Series A (DCVC, 2018), Series B (DCVC + Playground Global, 2020), Series C (DCVC, 2022, $150M), Extension/D (Amazon Industrial Innovation Fund, 2024, $150M+). Cumulative $500M+, estimated valuation ~$1.4B.
RoboFab manufacturing facility, Salem, Oregon — opened September 2023 as the first dedicated humanoid robot factory in the US. 70,000 sq ft. Stated 10,000 units/year production capacity. Source: companies.ts:agility-digit.strengths.
Knowledge-graph wiring per graph.ts: agility → actuators, nvidia-thor, force-torque, batteries-liion; investor edges amazon-iif → agility, dcvc → agility.
Component-level spec gaps reflect public-record absence as of writing. No production-unit teardown has been published by Munro & Associates, Morgan Stanley, China Merchants Securities, or equivalent third parties. Specific actuator model numbers, battery cell vendors, and contract manufacturers are not on the record.