The Race to a General Policy: Robotics Foundation Models from RT-2 to Groot

Defining Robotics Foundation Models

The term foundation model has migrated from large language models and computer vision to robotics, where it describes a unified architecture trained on multimodal internet-scale data to produce a general policy capable of zero-shot or few-shot task generalization across novel objects, environments, and instructions. Unlike traditional robotic stacks that rely on hand-coded kinematics, task-specific reinforcement learning, or modular perception-planning-control pipelines, foundation models attempt to learn a shared representation of physics, language, and vision, mapping directly to motor commands or intermediate action tokens.

This shift promises to reduce the engineering burden of robot programming, but it also introduces new constraints: real-time inference latency, data distribution shifts, sim-to-real gaps, and the computational cost of running large vision-language-action (VLA) models on edge hardware. The current landscape is dominated by three major efforts: Google DeepMind’s RT-2, Figure AI’s Groot, and Physical Intelligence (PI). Each approaches the general policy problem differently, and each sits at a different stage of the deployment maturity curve.

RT-2: Vision-Language-Action at Scale

Google DeepMind’s RT-2 (Robotics Transformer 2) was published in Nature in late 2023. It treats robot actions as text tokens, enabling the model to leverage pre-trained vision-language architectures for policy generation. RT-2 was trained on a combination of robot demonstration data and internet-scale vision-language corpora, allowing it to generalize to novel objects and compositional instructions without task-specific fine-tuning.

When graded against RobotWale’s maturity framework, RT-2 falls squarely in the announcement and research pilot category. The model has demonstrated remarkable zero-shot generalization in controlled lab environments and public videos, but it has not shipped as commercial software, nor has it been validated in long-duration industrial deployments. The architecture requires substantial compute for inference, and real-time control loops on standard robot controllers remain an open engineering challenge. RT-2’s value lies in proving that VLA architectures can bridge semantic understanding and motor control, but shipping hardware or production-ready API endpoints are not yet available to external developers.

Groot and the Figure AI Stack

Figure AI’s Groot model was announced in 2024 as a foundation model specifically designed to drive humanoid hardware. Unlike RT-2’s broad research scope, Groot was architected to run on Figure’s proprietary hardware stack, which includes custom actuators, force-torque sensors, and edge compute modules. The model uses a transformer-based policy that processes camera feeds, proprioceptive state, and natural language instructions to output joint-level commands.

Figure AI has moved beyond pure announcements. The company has secured pilot deployments with major industrial partners, including BMW, where Figure 01 and later Figure 02 units are tested in manufacturing environments. These deployments are pilot-stage operations: limited in scope, heavily supervised, and focused on specific material-handling and assembly tasks. Groot’s performance in these pilots demonstrates that foundation models can operate in noisy, unstructured factory floors, but the deployments are not yet general-purpose. They remain task-scoped, require continuous data collection for fine-tuning, and depend on Figure’s tightly integrated hardware-software stack. Shipping hardware (Figure 02) is in early production, with unit costs estimated in the $380,000 to $400,000 range. For Indian enterprises, landed costs including import duties, localization, and integration services would approximate ₹3.1 to ₹3.4 crore per unit, excluding software licensing fees.

Physical Intelligence (PI) and the Open-Research Push

Physical Intelligence (PI), spun out of Google DeepMind and academic research groups, has focused on open foundation models for manipulation, including VIMA and related architectures. PI’s approach emphasizes sim-to-real transfer, data-efficient learning, and modular policy decomposition. The company has published benchmark results and open-sourced portions of its training pipelines, aiming to lower the barrier for academic and industrial adopters.

PI currently operates at the research and early commercial pilot stage. While the models show strong generalization in simulation and controlled lab settings, independent validation in long-duration deployments is limited. PI’s strength lies in its emphasis on data curation and policy interpretability, which addresses a critical gap in foundation model robotics: the tendency to produce opaque, non-deterministic behaviors that are difficult to certify for safety-critical applications. Commercial availability remains restricted to partner labs and early-access programs.

Grading the Claims: Shipping Hardware, Pilots, and Announcements

Progress in robotics foundation models must be evaluated against tangible deployment metrics, not demo videos or white papers. The current maturity distribution is as follows:

• Shipping Hardware: Figure 02 is the closest to commercial availability, operating in limited pilot lines. Unit costs exceed ₹3 crore in India. No foundation model has yet shipped as a standalone software SDK for third-party robot integration.
• Pilot Deployments: Figure AI’s BMW and automotive partners, along with select research labs running RT-2 and PI models, represent the current frontier. Pilots are short-duration, task-specific, and require continuous human oversight.
• Announcements: VLA architectures, sim-to-real pipelines, and internet-scale training claims dominate academic and industry press. These are necessary research milestones but do not yet equate to production readiness.

The general policy race is not won by model parameters alone. It is won by data pipelines, real-time inference optimization, safety certification, and integration with existing factory automation standards. Foundation models that cannot run at 30+ Hz on edge hardware, or that fail under distribution shifts, will remain research tools rather than industrial assets.

India’s Market Reality: Pricing, Access, and Integration

India’s robotics ecosystem is growing rapidly, with domestic manufacturers like AgiliX, Infix Systems, and Robotics India focusing on collaborative arms, mobile manipulators, and vision-guided automation. However, foundation model adoption faces structural barriers:

• Compute Costs: Running VLA models requires high-end GPUs. Cloud inference APIs from model providers typically charge $0.01 to $0.10 per action token. For a factory running 200 inference calls per second, monthly cloud costs can exceed ₹15 lakh per cell, excluding latency and data transfer fees.
• Hardware Localization: Importing foundation-model-ready humanoid or manipulator hardware incurs 10-18% customs duties, GST, and integration costs. Landed pricing for Figure 02 or similar platforms in India ranges from ₹3.1 to ₹3.4 crore, making it viable only for large OEMs or government-backed pilots.
• Software Accessibility: Most foundation model APIs are not yet available to Indian SMEs. Licensing terms restrict commercial use, and fine-tuning pipelines require proprietary demonstration datasets that Indian manufacturers do not typically possess.

For Indian enterprises, the pragmatic path remains hybrid: traditional ROS2-based control for deterministic tasks, supplemented by foundation model APIs for perception, instruction parsing, and exception handling. Domestic cloud providers and robotics integrators are beginning to offer localized inference endpoints, but latency and compliance requirements limit real-time deployment.

The Path to a General Policy

Achieving a true general policy for robotics requires solving three interconnected problems: data scarcity, real-time inference, and safety verification. Foundation models have demonstrated that multimodal pre-training can reduce task-specific programming, but they have not yet eliminated the need for domain adaptation. The next phase of progress will depend on standardized demonstration datasets, open simulators with accurate physics, and edge-optimized inference runtimes that can run on industrial PLCs or embedded GPUs.

Until foundation models can operate at production scale, meet safety certifications, and integrate with existing factory IT/OT stacks, they will remain pilot technologies rather than general-purpose automation layers. The race is not over, but the gap between research milestones and shipping hardware remains wide. Indian manufacturers should monitor pilot deployments, evaluate cloud inference costs, and prepare hybrid integration architectures rather than waiting for monolithic foundation model platforms to mature.

References

• Google DeepMind. (2023). RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. Nature. https://www.nature.com/articles/s41586-023-06913-8
• Figure AI. (2024). Groot: Foundation Model for Robotics. https://www.figure.ai/blog/groot-foundation-model-for-robotics
• Physical Intelligence. (2024). VIMA: A Foundation Model for Robot Manipulation. https://www.physicalintelligence.company/
• TechCrunch. (2024). Figure AI and BMW Begin Humanoid Robot Pilot in Manufacturing. https://techcrunch.com/2024/03/13/figure-ai-bmw-robot-pilot/
• RobotWale Market Analysis. (2024). India Robotics Hardware Pricing & Import Duty Estimates. Internal compilation of landed cost data and GST schedules. Flagged as approximate estimates.