Physical Intelligence Pi API Beta Speeds Figure 03 and Tiangong Fine-Tuning

Pi API Beta Opens Hosted Fine-Tuning for Robot Foundation Models

Physical Intelligence launched its Pi API beta on October 13, 2026, giving robotics teams a hosted path to upload demonstration data, adapt the company’s latest π0 and π0.5 foundation models on managed infrastructure, and deploy policies through a cloud inference gateway. The waitlist requires applicants to describe intended use cases, and early partners already report functional behaviors after just three to four days of iteration. This service directly addresses one of the largest cost and time bottlenecks in humanoid and dual-arm deployments: the engineering hours required to move a generalist model onto specific hardware and tasks.

The announcement positions Physical Intelligence as infrastructure rather than a robot maker. Teams supply their own robot data and evaluation loops while the company handles compute, model updates, and inference endpoints. Early results include Almond’s Axol dual-arm system learning to deburr assembly spacers in four days and General Instinct reaching parity with an existing adapted open model using only 5 percent of the prior data volume. These numbers matter because fine-tuning cycles have historically stretched weeks or months when teams maintain their own training clusters and data pipelines.

Impact on Figure 03 Fleet Economics and Production Ramps

Figure AI has already produced more than 1,000 Figure 03 units through its BotQ line as of July 2026, with production rates reaching one robot per hour. The company operates logistics sequencing pilots at BMW’s Spartanburg plant in Hall 52 and maintains partner-only home trials. Each additional behavior or environment adaptation currently demands internal data collection and model work. Access to the Pi API could compress those cycles, lowering the per-robot engineering cost that scales linearly with fleet size.

At current production velocity, even modest reductions in adaptation time translate into earlier revenue recognition for Figure. A four-day iteration loop versus a three-week internal cycle means more behaviors can be validated before a new batch ships. The hosted inference option also reduces on-robot compute requirements, which directly affects bill-of-materials cost and power budget—two variables that matter when leasing robots at the $400–600 monthly range discussed by Figure leadership. Reliability metrics such as mean time between failures improve when policies are refined against diverse partner data rather than a single team’s test fleet.

Tiangong Ultra and Broader Chinese Humanoid Programs

The Beijing Humanoid Robot Innovation Center’s Tiangong Ultra has demonstrated sprint times under nine seconds and won multiple medals at the 2026 World Humanoid Robot Games, yet verified customer deployments remain limited. Mass delivery of the Tiangong 3.0 platform is slated for the second half of 2026. The Pi API’s emphasis on rapid fine-tuning for agriculture, server assembly, and material finishing tasks aligns with the use cases Chinese programs target for production-line and logistics work.

If Tiangong teams adopt the service, they gain access to models already stress-tested across more than a dozen partner robots without building equivalent training infrastructure. This matters for fleet economics because Chinese manufacturers accounted for 97 percent of global humanoid shipments in the first half of 2026. Any reduction in per-unit software development cost improves gross margins at scale. The API’s knowledge-insulation training method in π0.5 could also help preserve performance when models are updated mid-deployment, a practical concern when hundreds of units operate across multiple sites.

Technical Workflow and Second-Order Cost Implications

The beta workflow follows three steps: upload robot demonstration data, launch fine-tuning on Physical Intelligence-managed compute, then query the resulting policy through the hosted gateway. Evaluation data collected on the customer robot feeds the next iteration. The company states it will continue pushing its newest models through the same API, so partners receive capability improvements without re-architecting their own training stacks.

For operators running mixed fleets, the service creates an interesting economic lever. Instead of maintaining separate ML teams for each robot platform, a single fine-tuning contract can target Almond Axol, Figure 03, or Tiangong hardware. This shared layer reduces duplicated engineering headcount. It also concentrates failure-mode data across more hardware variants, potentially accelerating robustness gains that individual teams would reach more slowly.

Serviceability implications follow. When inference runs in the cloud, on-robot hardware stays simpler and field-replaceable units become fewer. Network latency and connectivity become new variables, however, and the announcement notes that some early applications encountered connectivity challenges. Operators must weigh those trade-offs against the capital and talent savings of outsourcing the model stack.

Competitive Positioning Among Humanoid Programs

Tesla Optimus, Boston Dynamics Atlas Electric, Unitree H1 and G1, and Agility Digit all face the same adaptation bottleneck. The Pi API offers a third-party route that does not require licensing a full robot platform. Figure AI’s own Helix models and Beijing’s WISE KAIWU foundation represent vertical integration strategies. Physical Intelligence instead sells the intelligence layer as a service, betting that many hardware makers will prefer to focus on mechanics, actuators, and end-effectors.

Early partner diversity—autonomous heavy equipment, food service, dual-arm manufacturing, and server assembly—suggests the models already generalize across payload and precision regimes relevant to both Figure 03 and Tiangong. Whether those models deliver comparable zero-shot performance in unseen homes or factories will determine uptake. The four-day Al