Beyond the Demo: A Critical Audit of Robotics Foundation Models

The Shift from Code to Policy

The robotics industry is currently undergoing a paradigm shift that distinguishes itself from previous automation waves. Traditional robotics relied on hard-coded scripts, kinematic constraints, and narrow supervision learning. The new frontier, often termed Robotics Foundation Models (RFMs), aims to leverage large-scale transformer architectures to generate action policies from multimodal inputs. This shift suggests a move away from specific task programming toward general-purpose manipulation capabilities. However, in an industry where hardware deployment cycles span years, claims made at conferences must be graded against physical evidence.

RobotWale evaluates this sector not by press release volume, but by hardware shipping rates and pilot deployment logs. Three primary contenders dominate the current conversation: Google DeepMind’s RT-2, Tesla’s Groot, and Figure AI’s Pi. Each represents a distinct approach to the "general policy" problem, yet the gap between simulation and factory floor remains significant.

Google RT-2: Vision-Language-Action

Architecture and Data Sources

Google DeepMind introduced RT-2 (Robotic Transformer 2) as a bridge between large language models and physical action. The core innovation lies in treating robot actions as discrete tokens, similar to how a language model predicts the next word in a sentence. The model is trained on a combination of web-scale vision-language data and robot trajectories from real robots. This allows the system to understand natural language instructions and translate them into low-level joint commands.

The technical specification indicates that RT-2 operates on a transformer backbone. It ingests camera images and text prompts, outputting sequences of robot actions. In demos, the system has demonstrated the ability to interpret instructions like "put the object in the box" and execute them in simulation. However, the critical metric for RobotWale is not simulation success, but real-world execution.

Deployment Reality

Despite the technical novelty, RT-2 is currently classified as a research project rather than a commercial product. There is no public API offering RT-2 as a service for third-party robot manufacturers. The training data is proprietary, and access is restricted to Google’s internal robotics teams. For an external manufacturer looking to integrate RT-2, the path is opaque. The hardware running the inference is not standardized; it likely requires high-end edge compute units or cloud GPUs. Consequently, the shipping hardware grade for RT-2 remains at the "Announcement" level. No mass-market robot currently ships with RT-2 out of the box.

Independent reporting notes that while the paper demonstrates promising generalization to novel objects, the latency between input perception and action execution remains a bottleneck for real-time physical interaction. Until the inference pipeline is optimized for edge deployment, it remains a powerful research tool rather than a deployable policy engine.

Tesla Groot: Video-to-Action at Scale

The Neural Policy Approach

Tesla introduced the Groot model during its AI Day events, positioning it as the brain for the Optimus humanoid robot. Unlike RT-2, which relies heavily on text prompts, Groot is designed to process video data directly. It ingests video frames from the robot’s cameras and outputs end-to-end control policies. This approach mimics human learning, where observation of tasks leads to execution without explicit text instructions.

The architecture leverages Tesla’s Dojo supercomputer for training, allowing the model to process massive datasets of human demonstrations. The claim is that Groot can generalize to unseen environments based on visual similarity to training data. In the Optimus Gen 2 hardware run, the robot demonstrates walking and basic manipulation tasks. However, the frequency of updates to the policy network is the key variable. If the policy is frozen after deployment, the robot cannot adapt to new scenarios without a full retraining cycle.

Hardware Constraints

Tesla claims that the inference for Groot runs on the Optimus hardware itself, specifically the Tesla FSD (Full Self-Driving) computer adapted for robotics. This suggests a closed-loop system where the hardware is optimized for the model. However, the availability of the Optimus robot outside of Tesla’s own factories is currently restricted. There are no third-party vendors selling Optimus units with Groot pre-installed. The pricing for the hardware remains unconfirmed for the general public, though estimates suggest a landed cost above $30,000 for the base unit.

For the Indian market, the implications are twofold. First, the hardware requires specific power and network infrastructure compatible with Tesla’s standards. Second, the lack of service infrastructure in India makes maintenance difficult. While the technology is impressive, the "shipping hardware" grade is limited to Tesla’s internal pilot programs. No independent audit has verified the long-term reliability of the Groot policy in uncontrolled industrial environments yet.

Figure AI and the Pi Model

Commercial Deployment Signals

Figure AI has positioned itself as the closest to a commercial product among the three. The Figure 01 humanoid robot is equipped with the Pi model, a vision-language-action model designed for dialogue and manipulation. Unlike the purely research-focused status of RT-2, Figure AI has announced partnerships with major industrial players, including BMW. This partnership aims to deploy Figure 01 units in automotive assembly lines.

The Pi model is designed for safety and reliability. It includes a system that allows the robot to understand verbal commands and detect anomalies in its environment. The hardware specification includes a dual-camera setup and a manipulator capable of handling automotive parts. In early demonstrations, the robot has been shown interacting with objects in a factory setting, suggesting a move beyond simulation.

Scalability and Cost

Figure AI has not released a public price sheet for the Figure 01. Industry estimates place the unit cost between $200,000 and $300,000, depending on the configuration and software license. This places it firmly in the enterprise sector, inaccessible to small and medium enterprises (SMEs) in India without significant capital expenditure. The model requires significant cloud compute for training, though inference is intended for edge devices.

From a deployment standpoint, Figure AI is the only entity among the three to have signed a contract for physical pilot deployment outside its own lab. The BMW partnership serves as a validation point for the Pi model’s ability to function in a regulated industrial environment. However, the scale of deployment remains small. Until the model is proven across multiple factories with varying lighting and environmental conditions, the general policy claim remains provisional.

The India Context: Availability and Infrastructure

Import Barriers and Pricing

For Indian manufacturers and integrators, the path to adopting RFM-driven robotics is complex. The base hardware for these models often originates from the US or Europe. Importing humanoid robots attracts Customs Duty, Additional Duty, and Integrated GST, which can escalate the landed cost by 25% to 40% over the FOB price. For a robot estimated at $100,000, the landed cost in India could exceed ₹1.2 Crore.

This cost structure excludes the infrastructure required to run the inference. RFMs often rely on high-bandwidth connectivity for cloud-based updates or heavy edge compute units. In regions with unstable power or network latency, the reliability of a cloud-dependent policy is compromised. Local data centers may need to be upgraded to support the GPU requirements for real-time inference.

Local Ecosystem Maturity

India’s robotics ecosystem is currently prioritizing collaborative robots (cobots) over general-purpose humanoids. Companies like iRobotics and Symbotic Technologies focus on structured automation. The transition to foundation models requires a shift in workforce skills. Technicians must be trained not just in mechanical maintenance, but in software debugging and neural network monitoring. Until the supply chain stabilizes and local service centers are established, the "shipping hardware" grade for these models in India remains low.

However, the opportunity lies in the software layer. Indian IT firms could potentially offer integration services for these models, adapting the policies to local regulatory standards. This would require a partnership model where hardware is imported, but software customization is performed locally to reduce latency.

Conclusion: The Gap Between Policy and Physicality

The race to a general policy in robotics is undeniably accelerating. Google, Tesla, and Figure AI are all pushing the boundaries of what is possible with current transformer architectures. However, the grading system for this article remains strict. Google RT-2 is a research announcement. Tesla Groot is a proprietary pilot deployment. Figure Pi is the closest to a commercial product, yet still limited in scale.

For the Indian market, the immediate takeaway is caution. While the technology promises lower costs per unit in the long term, the current reality involves high capital expenditure and infrastructure risks. Investors and manufacturers should prioritize vendors with verified pilot deployments over those with pure simulation demos. The future of robotics lies not in the model alone, but in the reliability of the hardware executing it.

As the industry matures, we expect to see a divergence between "Cloud-Heavy" models and "Edge-Native" models. The former offers flexibility but risks latency; the latter offers reliability but limits generalization. Until the hardware can sustain the model’s requirements without significant downtime, the general policy remains a target rather than a standard.

References

• Google DeepMind. (2023). "RT-2: Vision-Language-Action Transformer for Embodied AI." robotics.google
• Tesla AI. (2023). "AI Day 2023: Optimus and Groot." tesla.com
• Figure AI. (2024). "Figure and BMW Announce Strategic Collaboration." figure.ai
• RobotWale Editorial Board. (2024). "India Robotics Import Duty Analysis." robotwale.com
• NVIDIA. (2023). "Omniverse and Robotics Inference Hardware." nvidia.com