Chapter 9: US Foundation-Model Labs: PI, Generalist, Skild, and Figure

Overview

PI, Generalist, Skild, and Figure all speak in generalist terms, yet they differ in data source, embodiment, action interface, contact channel, and production evidence. This chapter rewrites the problem as a factory-cell data contract. The promise of large robot datasets in ^[1] and ^[2] matters, but in manufacturing an episode becomes valuable only when it is attached to inspection and process outcomes.

PI has comparatively strong research openness through pi0 and OpenPI; Figure has strong humanoid hardware and company disclosures; Generalist and Skild expose fewer technical details. They must be compared with evidence tiers. The data is therefore not just a camera stream. It includes hand choice, force-torque, tactile patches, controller mode, operator intervention, and inspection result. ^[3] and ^[4] show why human-to-robot transfer loses critical signals unless the capture interface is designed around embodiment and contact.

After reading this chapter... - Explain factory manipulation as a data contract rather than a model score. - Separate two-finger, suction, custom, and five-finger choices by collection cost and failure observability. - Read papers, company releases, and deployment claims through evidence tiers. - Design the replay set and QA trace required for a first manufacturing PoC.

Core Map

Visual Argument

Figure 9.1. Company landscape for large-data manipulation. Source: reused local survey asset or author-created illustration.

Figure 9.2. Figure Helix as company-reported humanoid VLA stack. Source: reused local survey asset or author-created illustration.

Figure 9.3. pi0 family as a generalist policy reference point. Source: reused local survey asset or author-created illustration.

Decomposing the Word Generalist

The practical meaning of decomposing the word generalist is that process variables must be left in a learnable form. ^[1] shows how scale can broaden embodiment and task families, yet a factory cell is narrower and stricter. A pick that looks identical to a person can become a different distribution when fixture tolerance, surface contamination, cycle pressure, or reject code changes.

In the concrete scenario, PI has comparatively strong research openness through pi0 and OpenPI; Figure has strong humanoid hardware and company disclosures; Generalist and Skild expose fewer technical details. They must be compared with evidence tiers. Human demonstrations alone are not enough. Even when an interface such as ^[2] makes collection easier, a policy will repeat the same release failure if contact force and failure cause are absent. The episode schema must bind observation, action, contact state, QA outcome, and operator note under one key.

The evidence tier matters. A source such as ^[3] gives a method and benchmark that can be inspected. A company source can reveal deployment direction, but usually exposes less about data rights, recovery handling, and operating metrics. Manufacturers should place both in the same comparison table but never read them with the same confidence.

From a control perspective, the learned policy should not be the whole system. Force limits, guarded motion, fixture state, collision zones, and rollback conditions have to sit around the model. Without that boundary, collecting more data can create more safety stops and more rework instead of a more deployable robot.

From an operating perspective, the reproducibility of failure matters more than a headline success rate. A failure has to enter a replay set, the next update has to pass that replay set, and the deployed cell has to show a lower rate for the same failure code. The question is not only how to describe decomposing the word generalist, but which missing log would make the claim collapse on a line.

PI and OpenPI Openness

The practical meaning of pi and openpi openness is that process variables must be left in a learnable form. ^[4] shows how scale can broaden embodiment and task families, yet a factory cell is narrower and stricter. A pick that looks identical to a person can become a different distribution when fixture tolerance, surface contamination, cycle pressure, or reject code changes.

In the concrete scenario, PI has comparatively strong research openness through pi0 and OpenPI; Figure has strong humanoid hardware and company disclosures; Generalist and Skild expose fewer technical details. They must be compared with evidence tiers. Human demonstrations alone are not enough. Even when an interface such as ^[5] makes collection easier, a policy will repeat the same release failure if contact force and failure cause are absent. The episode schema must bind observation, action, contact state, QA outcome, and operator note under one key.

The evidence tier matters. A source such as ^[6] gives a method and benchmark that can be inspected. A company source can reveal deployment direction, but usually exposes less about data rights, recovery handling, and operating metrics. Manufacturers should place both in the same comparison table but never read them with the same confidence.

From a control perspective, the learned policy should not be the whole system. Force limits, guarded motion, fixture state, collision zones, and rollback conditions have to sit around the model. Without that boundary, collecting more data can create more safety stops and more rework instead of a more deployable robot.

From an operating perspective, the reproducibility of failure matters more than a headline success rate. A failure has to enter a replay set, the next update has to pass that replay set, and the deployed cell has to show a lower rate for the same failure code. The question is not only how to describe pi and openpi openness, but which missing log would make the claim collapse on a line.

Figure's Humanoid-Stack Strategy

The practical meaning of figure's humanoid-stack strategy is that process variables must be left in a learnable form. ^[7] shows how scale can broaden embodiment and task families, yet a factory cell is narrower and stricter. A pick that looks identical to a person can become a different distribution when fixture tolerance, surface contamination, cycle pressure, or reject code changes.

In the concrete scenario, PI has comparatively strong research openness through pi0 and OpenPI; Figure has strong humanoid hardware and company disclosures; Generalist and Skild expose fewer technical details. They must be compared with evidence tiers. Human demonstrations alone are not enough. Even when an interface such as ^[8] makes collection easier, a policy will repeat the same release failure if contact force and failure cause are absent. The episode schema must bind observation, action, contact state, QA outcome, and operator note under one key.

The evidence tier matters. A source such as ^[9] gives a method and benchmark that can be inspected. A company source can reveal deployment direction, but usually exposes less about data rights, recovery handling, and operating metrics. Manufacturers should place both in the same comparison table but never read them with the same confidence.

From a control perspective, the learned policy should not be the whole system. Force limits, guarded motion, fixture state, collision zones, and rollback conditions have to sit around the model. Without that boundary, collecting more data can create more safety stops and more rework instead of a more deployable robot.

From an operating perspective, the reproducibility of failure matters more than a headline success rate. A failure has to enter a replay set, the next update has to pass that replay set, and the deployed cell has to show a lower rate for the same failure code. The question is not only how to describe figure's humanoid-stack strategy, but which missing log would make the claim collapse on a line.

Uncertainty Around Generalist and Skild

The practical meaning of uncertainty around generalist and skild is that process variables must be left in a learnable form. ^[10] shows how scale can broaden embodiment and task families, yet a factory cell is narrower and stricter. A pick that looks identical to a person can become a different distribution when fixture tolerance, surface contamination, cycle pressure, or reject code changes.

In the concrete scenario, PI has comparatively strong research openness through pi0 and OpenPI; Figure has strong humanoid hardware and company disclosures; Generalist and Skild expose fewer technical details. They must be compared with evidence tiers. Human demonstrations alone are not enough. Even when an interface such as ^[11] makes collection easier, a policy will repeat the same release failure if contact force and failure cause are absent. The episode schema must bind observation, action, contact state, QA outcome, and operator note under one key.

The evidence tier matters. A source such as ^[12] gives a method and benchmark that can be inspected. A company source can reveal deployment direction, but usually exposes less about data rights, recovery handling, and operating metrics. Manufacturers should place both in the same comparison table but never read them with the same confidence.

From a control perspective, the learned policy should not be the whole system. Force limits, guarded motion, fixture state, collision zones, and rollback conditions have to sit around the model. Without that boundary, collecting more data can create more safety stops and more rework instead of a more deployable robot.

From an operating perspective, the reproducibility of failure matters more than a headline success rate. A failure has to enter a replay set, the next update has to pass that replay set, and the deployed cell has to show a lower rate for the same failure code. The question is not only how to describe uncertainty around generalist and skild, but which missing log would make the claim collapse on a line.

Evidence Tiers for Manufacturers

The practical meaning of evidence tiers for manufacturers is that process variables must be left in a learnable form. ^[13] shows how scale can broaden embodiment and task families, yet a factory cell is narrower and stricter. A pick that looks identical to a person can become a different distribution when fixture tolerance, surface contamination, cycle pressure, or reject code changes.

In the concrete scenario, PI has comparatively strong research openness through pi0 and OpenPI; Figure has strong humanoid hardware and company disclosures; Generalist and Skild expose fewer technical details. They must be compared with evidence tiers. Human demonstrations alone are not enough. Even when an interface such as ^[14] makes collection easier, a policy will repeat the same release failure if contact force and failure cause are absent. The episode schema must bind observation, action, contact state, QA outcome, and operator note under one key.

The evidence tier matters. A source such as ^[15] gives a method and benchmark that can be inspected. A company source can reveal deployment direction, but usually exposes less about data rights, recovery handling, and operating metrics. Manufacturers should place both in the same comparison table but never read them with the same confidence.

From a control perspective, the learned policy should not be the whole system. Force limits, guarded motion, fixture state, collision zones, and rollback conditions have to sit around the model. Without that boundary, collecting more data can create more safety stops and more rework instead of a more deployable robot.

From an operating perspective, the reproducibility of failure matters more than a headline success rate. A failure has to enter a replay set, the next update has to pass that replay set, and the deployed cell has to show a lower rate for the same failure code. The question is not only how to describe evidence tiers for manufacturers, but which missing log would make the claim collapse on a line.

Manufacturing Cell Checkpoint

Start a PoC by writing the data contract before selecting the model. PI has comparatively strong research openness through pi0 and OpenPI; Figure has strong humanoid hardware and company disclosures; Generalist and Skild expose fewer technical details. They must be compared with evidence tiers. The contract should include episode ID, operator or teleop ID, hand or tool ID, contact channel, controller mode, quality decision, defect code, override reason, and rollback condition. If a vendor hides or cannot export these fields, the manufacturer cannot explain why model improvement happens.

The second checkpoint is evidence tiering. Papers with arXiv or DOI links support method claims; official company pages support product-direction claims; press and watchlist items should not become load-bearing claims. PI, Generalist, Skild, Figure, Covariant, Dexterity, Chef Robotics, Sunday, Config, and CarbonSix all fit under data-driven manipulation, but they expose very different evidence about manufacturing readiness.

Open Questions and Failure Modes

First, large data without contact state leaves insertion, wiping, deformable handling, and tool use under-observed. Second, when data rights stay entirely with the vendor, the manufacturer does not own the cause of process improvement. Third, worker video and process IP are learning assets and governance risks at the same time. Fourth, sim-to-real data does not guarantee factory performance unless it is tied to real QA labels.

Data Contract Addendum

The important unit is not the model name but the observable attempt. To a human, two factory cycles can look like the same task; to a robot, they may differ in surface friction, initial pose, contact order, force limit, and inspection rule. Data strategy is therefore an operating system for connecting failures to process variables, not a campaign for collecting success clips.

Factory data is slower and messier than benchmark data, but it carries more decisive labels. Defect codes, rework, operator intervention, line stops, and safety stops are not noise to be hidden from learning. They are the labels that decide whether a policy can be deployed. If those labels are detached from episodes, model curves can improve while process KPIs stay flat.

Large-data driven manipulation is therefore not a choice of one VLA. It is a joint design of task schema, hand choice, controller boundary, replay set, QA trace, and update governance. Without that design, a foundation model may produce impressive demonstrations without becoming repeatable replacement labor in manufacturing.

Operating Loop Addendum

The important unit is not the model name but the observable attempt. To a human, two factory cycles can look like the same task; to a robot, they may differ in surface friction, initial pose, contact order, force limit, and inspection rule. Data strategy is therefore an operating system for connecting failures to process variables, not a campaign for collecting success clips.

Factory data is slower and messier than benchmark data, but it carries more decisive labels. Defect codes, rework, operator intervention, line stops, and safety stops are not noise to be hidden from learning. They are the labels that decide whether a policy can be deployed. If those labels are detached from episodes, model curves can improve while process KPIs stay flat.

Large-data driven manipulation is therefore not a choice of one VLA. It is a joint design of task schema, hand choice, controller boundary, replay set, QA trace, and update governance. Without that design, a foundation model may produce impressive demonstrations without becoming repeatable replacement labor in manufacturing.

Deployment Governance Addendum

The important unit is not the model name but the observable attempt. To a human, two factory cycles can look like the same task; to a robot, they may differ in surface friction, initial pose, contact order, force limit, and inspection rule. Data strategy is therefore an operating system for connecting failures to process variables, not a campaign for collecting success clips.

Factory data is slower and messier than benchmark data, but it carries more decisive labels. Defect codes, rework, operator intervention, line stops, and safety stops are not noise to be hidden from learning. They are the labels that decide whether a policy can be deployed. If those labels are detached from episodes, model curves can improve while process KPIs stay flat.

Large-data driven manipulation is therefore not a choice of one VLA. It is a joint design of task schema, hand choice, controller boundary, replay set, QA trace, and update governance. Without that design, a foundation model may produce impressive demonstrations without becoming repeatable replacement labor in manufacturing.

What to Learn Next

The next chapter applies the same principle to hands and end-effectors. Some cells are best served by suction or two fingers; others cannot close the data loop without five-finger hands and tactile or force-rich data.

References

1. Black, Kevin (2024). pi0: A Vision-Language-Action Flow Model for General Robot Control. arXiv.
2. Physical Intelligence (2024). pi0: A Generalist Robot Policy. Company research post.
3. Physical Intelligence (2025). OpenPI: Open Source Robot Policy Stack. GitHub.
4. Kim, Moo Jin (2024). OpenVLA: An Open-Source Vision-Language-Action Model. arXiv.
5. Octo Model Team (2024). Octo: An Open-Source Generalist Robot Policy. arXiv.
6. O'Neill, Abby (2023). Open X-Embodiment: Robotic Learning Datasets and RT-X Models. arXiv.
7. Pertsch, Karl (2025). FAST: Efficient Action Tokenization for Vision-Language-Action Models. arXiv.
8. Shukla, Shivin (2025). SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics. arXiv.
9. DeepMind Robotics Team (2025). Gemini Robotics: Bringing AI into the Physical World. arXiv.
10. Bjorck, Johan (2025). GR00T N1: An Open Foundation Model for Generalist Humanoid Robots. arXiv.
11. NVIDIA (2025). Isaac GR00T N1 Open Foundation Model for Humanoid Robots. NVIDIA Developer.
12. Figure AI (2025). Helix: A Vision-Language-Action Model for Generalist Humanoid Control. Company announcement.
13. Figure AI (2026). Figure 03 + Helix 02: General-Purpose Humanoid System. Company product page.
14. Generalist AI (2025). GEN-0 Robot Foundation Model. Company page.
15. Skild AI (2024). General-Purpose Robot Brain. Company page.
16. Zheng, Renhao (2026). EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data. NVIDIA Research.
17. Yu, Wenhao (2025). ForceVLA: Enhancing VLA Models with a Force-aware MoE for Contact-rich Manipulation. arXiv.
18. Huang, Yuhang (2025). Tactile-VLA: Unlocking Vision-Language-Action Model's Physical Knowledge for Tactile Generalization. arXiv.
19. Khazatsky, Alexander (2024). DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset. arXiv.
20. AgiBot World Team (2025). AgiBot World Colosseo: A Large-Scale Manipulation Platform. arXiv.
21. Zhao, Tony Z. (2023). Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware. arXiv.
22. Chi, Cheng (2023). Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. arXiv.
23. Ha, Huy (2024). Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots. arXiv.
24. Toyota Research Institute (2024). Large Behavior Models for Robot Manipulation. Company technical post.