Robotics Foundation Models: Assessing Pi, RT-2, and the General Policy Race

The Reality of Robotics Foundation Models

The term "foundation model" has migrated from large language models (LLMs) to robotics, promising a shift from task-specific programming to general-purpose agents. However, the editorial stance of RobotWale.com remains grounded: claims are graded by shipping hardware first, pilot deployments second, and announcements last. While the theoretical framework suggests that a single model could control diverse robot bodies across industries, the current landscape reveals a fragmented reality where software capabilities often outpace hardware reliability.

This article evaluates three significant entrants in the race for general robotics policy: Google DeepMind's RT-2, Tesla's Groot, and Figure AI's Pi. The focus is strictly on demonstrated capabilities, available documentation, and deployment status rather than marketing slides.

Google DeepMind: RT-2 and Data-Centric Vision

Google DeepMind's RT-2 (Robots Transformed) represents a significant shift in how robots interpret the world. Unlike traditional robotic control stacks that rely on explicitly programmed trajectories, RT-2 treats robotics as a language problem. It maps visual inputs and language commands directly to robot actions using a transformer architecture trained on massive datasets of internet images and robotic trajectories.

According to the DeepMind technical report, the model leverages a dataset of over 200,000 robotic demonstrations combined with internet-scale vision-language data. The system allows a robot to understand phrases like "pick up the red screwdriver" and generalize to objects it has not seen before during training, provided they exist in the training distribution.

However, the deployment reality is constrained. As of late 2023 and early 2024, RT-2 has primarily been demonstrated in simulation or on specialized hardware setups like the Google RT-2 robot arm. It is not yet a commercially available package that can be installed on a third-party humanoid. The model requires specific actuation interfaces to translate language outputs into motor commands.

Key technical constraints include:

• Latency: Inference on large vision transformers is computationally expensive, often requiring cloud processing which introduces latency unsuitable for real-time balance control.
• Hardware Dependency: The model does not run natively on edge hardware without significant optimization. Most current demos rely on high-end GPUs for inference.
• Generalization: While the model claims to handle novel objects, failure modes in physical environments remain unquantified in public reports.

For Indian manufacturers, the RT-2 architecture offers a research pathway but does not currently offer an off-the-shelf software solution. The cost of licensing or running such a model on-premise would likely exceed INR 50 lakhs for a single pilot deployment due to compute requirements.

Tesla Optimus and the Groot Model

Tesla's entry into the robotics space focuses heavily on the hardware first, with the software (Groot) serving as the control layer. Tesla's approach differs from DeepMind by prioritizing the end-to-end neural network training on actual robot data rather than internet datasets.

The Groot model is trained on video data captured by the Optimus robots themselves, utilizing Tesla's Dojo supercomputing infrastructure. The claim is that by ingesting human demonstrations and robot self-correction data, the network learns natural movement policies. Video demonstrations from the Tesla AI Day events show the Optimus Bot performing tasks like walking, carrying boxes, and sorting items.

Hardware Status:

• Shipping Hardware: Tesla has announced the delivery of early prototypes to employees for internal testing. As of early 2024, no commercial units are available for purchase.
• Performance: The hardware relies on a custom-designed electric powertrain and a control stack that integrates vision and actuation tightly.
• Cost: Elon Musk has cited a target price of under $20,000 USD for mass production. In India, landed costs including import duties, GST, and localization would likely push the initial price above INR 25 lakhs even before software licensing is considered.

It is critical to note that while the Groot model shows promise in simulation and limited field tests, the reliability of the hardware in unstructured environments remains the bottleneck. The "General Policy" is currently restricted to the specific kinematic chain of the Optimus unit. Porting this policy to other robot bodies requires significant retraining.

Figure AI: Pi and the BMW Pilot

Figure AI has gained attention through its partnership with BMW, specifically for manufacturing tasks. The Figure 01 robot, running the Pi model, has been demonstrated performing tasks such as sorting batteries and inspecting parts.

The Pi model utilizes a foundation model trained on a large dataset of robotic interactions, similar to RT-2 but with a focus on industrial safety and precision. The key distinction here is the commercial intent. Figure AI has signed a pilot deployment agreement with BMW Group, moving beyond the research phase into industrial validation.

Deployment Reality:

• Pilot Scale: The deployment is limited to specific BMW facilities. It is not a general release product for the Indian market.
• Hardware Specs: The robot features a torque-controlled actuation system designed for safety. It does not rely on reactive balance alone but on controlled actuation.
• Integration: The system requires significant integration with factory infrastructure (PLCs, safety sensors), which limits its "plug-and-play" appeal.

For the Indian automotive sector, the Figure Pi model presents a high-value opportunity but a high-barrier entry. The cost of a pilot unit is estimated at approximately $300,000 USD based on industrial robotics pricing benchmarks, translating to roughly INR 2.5 crores per unit. This excludes the software licensing fees which are likely enterprise-priced.

India Availability and Pricing Landscape

The availability of robotics foundation models in India is currently negligible for the general public and limited for enterprise clients. Most of these models are US-centric or Europe-centric due to the regulatory environment regarding safety and liability.

Import and Custom Duties: Importing a humanoid robot into India attracts a Base Customs Duty (BCD) of 10-15% plus an Additional Duty (Agriculture Infrastructure Development Cess) of 5%. With GST at 18%, the landed cost of a $20,000 unit effectively becomes approximately $28,000 USD (approx. INR 23 lakhs).

Service and Maintenance: Unlike consumer electronics, humanoid robots require specialized maintenance. Current Indian service infrastructure does not support the specific actuation systems used by Tesla or Figure AI. This creates a supply chain risk for any pilot deployment.

Software Licensing: Foundation models often operate under API access models or on-premise enterprise licenses. For a company like Tata Motors or Mahindra to deploy a robot, they would likely need to negotiate a contract that covers both hardware and the ongoing model updates.

The Path to General Policy

The race to a general policy is defined by one metric: shipping hardware. While the theoretical capability of models like RT-2 and Pi to understand natural language is impressive, the physical execution remains the primary constraint.

Grading the current state:

1. Shipping Hardware: Tesla Optimus and Figure 01 have prototypes in the field. Google RT-2 remains largely in research labs.
2. Pilot Deployments: Figure AI has a BMW pilot. Tesla has internal employee pilots. Google has limited external trials.
3. Announcements: Most claims of "general purpose" remain unverified by third-party auditors.

Until a model can be verified on a standard robotic body in a public environment, the "Foundation Model" label remains a research classification rather than a commercial product specification.

Conclusion

The robotics foundation model race is in its infancy. While the potential for a unified policy controlling multiple robot platforms is the ultimate goal, the current reality is a fragmented market where software is often tied to specific hardware constraints. For Indian stakeholders, the focus should be on hardware availability, service infrastructure, and landed costs rather than the theoretical capabilities of the models themselves.

RobotWale.com will continue to track these deployments, verifying claims against physical evidence. Until shipping hardware becomes the norm for these foundation models, they remain high-potential, high-risk propositions.