Frontline · Robotics · 2026-05-17

Figure demonstrates live package-sorting with Figure03 humanoid

The AI-first robotics company showed its third-generation humanoid picking, inspecting, and routing parcels in a live industrial workflow—the clearest signal yet that the platform is approaching warehouse-grade reliability.

Founded

2022

4 years

Status

Private

Total raised

$1.7B

Headcount

201-500

The story

Figure aired a live demonstration^[1] of its Figure03 humanoid executing package-sorting tasks in what appears to be a logistics environment—picking parcels from an inbound conveyor, inspecting barcodes or labels, and routing them to designated bins. The session ran without cuts or edits, a deliberate echo of the company's June 2025 60-minute endurance video but compressed into a workflow that mirrors the economic center of gravity for humanoid deployment: warehouse and fulfillment operations. The demo marks a step-function improvement in task complexity over prior exhibits, which focused on continuous operation rather than decision-dense manipulation. We're tracking this as the first time Figure has shown a humanoid performing a multi-step, perception-heavy logistics task in real time. The sorting workflow requires object recognition, grasp planning under variable lighting and occlusion, dynamic re-planning when packages arrive out of sequence, and enough dexterity margin to handle a mix of box sizes and weights without pre-staging. That's a materially higher bar than the static assembly tasks Figure demonstrated at BMW's South Carolina plant in early 2025, which involved fixed part geometry and structured handoffs. The shift signals that Figure's vision-language-action models—trained on human video as part of Project Go-Big^[1]—are beginning to generalize across unstructured environments, the long-promised unlock for commercial humanoid economics. The timing matters. Tesla Optimus is already operating inside Tesla's Fremont and Austin factories, and Agility Robotics' Digit is deployed at Amazon and third-party logistics providers. Figure raised over $1 billion in September 2025 with the explicit thesis that AI-native architectures would leapfrog incumbents who bolted language models onto classical control stacks. This demo is the first public proof that the thesis is landing: Figure03 isn't just durable, it's decision-capable at the tempo logistics operators require. The question now is cycle time—can it match or beat the human benchmark of 400–600 packages per hour that defines breakeven labor replacement in parcel sortation?

The rest of this story is for subscribers.

Including Our Take, the Tailwinds & headwinds framing, Connections across the FOBI roster, and What should you do.

Founding

50% off

/month

94 of 100 spots left

Full

$10

/month

Available once all 100 Founding Member spots are claimed.

Get full access

Already subscribed? Sign in →

In plain English

Figure is a company building human-shaped robots that can walk, pick things up, and do jobs in warehouses and factories. They just showed their newest robot, called Figure03, sorting packages in real time—moving boxes from one place to another, checking them, and putting them in the right spot. This is important because most robot demos are carefully staged; live sorting means the robot can handle messy, unpredictable work without breaking down.

Our Take

This demo is Figure's bid to redefine the humanoid value proposition from endurance spectacle to operational credibility. The June 2025 60-minute video proved Figure03 could run; this sorting workflow proves it can think under tempo. That distinction matters because the bottleneck in humanoid deployment has shifted from hardware reliability to decision latency—can the robot perceive, plan, and execute fast enough to match human cycle time in variable environments? Figure's choice to demo package sorting rather than structured assembly signals confidence that its AI-native architecture has crossed that threshold. The competitive implication: incumbents who treat foundation models as a feature bolted onto classical controls may find themselves structurally slower to generalize, even if their hardware is mature. The moat here isn't dexterity or runtime—it's the speed at which the platform learns new tasks from data.

Takeaways

01Figure03's live package-sorting demo is the first public evidence that AI-native humanoids can handle decision-dense, unstructured logistics tasks in real time.
02The workflow—perception, grasp planning, and dynamic routing—represents a step-function increase in complexity over Figure's prior BMW assembly demos.
03Figure's vision-language-action models, trained on human video via Project Go-Big, appear to be generalizing across variable environments—the economic unlock for humanoid labor replacement.
04The competitive benchmark is now cycle time: can Figure03 sort 400–600 packages per hour, the human baseline that defines breakeven economics in parcel operations?
05This positions Figure as a credible alternative to Tesla Optimus and Agility Robotics' Digit in the race to deploy general-purpose humanoids at warehouse scale.

Tailwinds & headwinds

Tailwinds

Logistics and fulfillment operators face structural labor shortages and peak-season capacity constraints, creating urgent demand for scalable humanoid automation.
AI-native architectures trained on human video can generalize across tasks faster than classical control systems, compressing time-to-deployment and reducing per-task engineering costs.
Figure's $1.7B in capital provides runway to iterate on hardware durability and cycle-time optimization without near-term revenue pressure.
The live, uncut demo format builds customer confidence in system reliability—critical for risk-averse warehouse operators evaluating humanoid pilots.

Headwinds

Cycle time remains unproven: if Figure03 can't match 400–600 packages per hour, labor-replacement economics collapse and humanoids become a premium solution for niche applications only.
Tesla Optimus can leverage vertical integration in AI, actuators, and manufacturing to undercut Figure on unit cost, compressing margins in a volume race.
Durability under 24/7 multi-shift operation is still unproven at scale—warehouse environments impose thermal, dust, and vibration stresses that lab demos don't capture.
Task-specific incumbents like AutoStore's cube-based AS/RS systems already deliver 99.9%+ uptime in sortation; humanoids must prove comparable reliability to displace installed…

What should you do

The asymmetric bet here is that AI-native architectures collapse the time-to-task-generalization curve faster than incumbents who retrofit foundation models onto classical control stacks. If Figure's vision-language-action models continue to scale with data—and this demo suggests they are—the company is positioned to capture high-margin logistics contracts ahead of Tesla Optimus's volume ramp and Agility Robotics' installed base. Capital flowing toward humanoid infrastructure plays—compute, actuator supply chains, training-data platforms—reflects the market's growing conviction that the economics of warehouse labor replacement are crossing breakeven. This demo strengthens that thesis. The positioning question is whether Figure can convert this proof of concept into multi-site deployments before Tesla l…

Strategic-positioning commentary · not investment advice

Data snapshot

Figure funding to date: $1.7B
Human sorter baseline (packages/hour): 400–600
Figure03 public cycle-time benchmark: undisclosed
Months since Project Go-Big launch: ~8
Prior Figure demo (June 2025 endurance test): 60 min continuous

Dependencies & bottlenecks

Training-data volume and diversity: Project Go-Big's human-video corpus must cover enough task variation to generalize across logistics workflows without per-site retraining.
Compute scaling for real-time inference: vision-language-action models require edge or near-edge GPU capacity to hit sub-second decision latency at warehouse tempo.
Actuator and drivetrain durability under 24/7 multi-shift operation—mechanical wear in joints and gearboxes remains the Achilles' heel for humanoid uptime in industrial environments.
Customer willingness to pilot unproven platforms: logistics operators are conservative; Figure needs anchor deployments to generate fleet-scale purchase orders.

Trajectory

Forward signals

Figure's first public disclosure of Figure03 cycle-time benchmarks in packages per hour, expected in pilot-customer case studies or technical publications by Q3 2026.
Multi-site logistics deployments beyond BMW—particularly partnerships with third-party logistics providers or e-commerce fulfillment networks—as proof of commercial scalability.
Tesla Optimus deployment announcements outside Tesla's own factories, which would signal Tesla's intent to compete head-to-head with Figure in external logistics contracts.
Durability and uptime data from 24/7 multi-shift operation in peak-season sortation, the real-world stress test that will validate or break Figure's labor-replacement thesis.

Report Notes

1.
CatalystDROIDS!
droids.substack.com · 2026-05-16

Cast

Figure
Tesla Optimus — vertically integrated humanoid rival
AutoStore — task-specific automation incumbent

Glossary

vision-language-action models: Neural networks that map visual input and natural-language instructions directly to robot motor commands, bypassing classical path-planning pipelines. Figure trains these models on human video to enable task generalization across unstructured environments.
cycle time: The elapsed time to complete one unit of work—in parcel sortation, the seconds required to pick, inspect, and route one package. Human sorters average 6–9 seconds per package (400–600 per hour); humanoids must match or beat this to justify labor-replacement economics.
grasp planning: The real-time computation of hand position, finger configuration, and force application required to securely grip an object under variable geometry, weight, and surface conditions. Critical for manipulation in unstructured logistics environments.
task generalization: A robot's ability to perform new tasks or handle novel object types without task-specific reprogramming. AI-native humanoids aim to generalize from training data; classical industrial robots require per-task engineering.
Project Go-Big: Figure's initiative, announced September 2025, to train humanoid control policies on large-scale human video datasets captured across Brookfield's global real estate portfolio. The goal is to teach robots generalizable manipulation skills by observing human behavior.