The Race to a General Policy: Foundation Models in Humanoid Robotics
The Foundation Model Shift in Robotics
The robotics industry is transitioning from hand-tuned, task-specific controllers to vision-language-action (VLA) architectures that generalize across environments. Foundation models in robotics are not large language models running on robot arms; they are multimodal neural networks trained on massive datasets of robot trajectories, sensory streams, and language instructions. Their output is a continuous policy that maps raw perception to low-level motor commands, enabling zero-shot or few-shot adaptation to novel objects, lighting conditions, and spatial layouts.
This shift is real, but the gap between research milestones and commercial deployment remains wide. At RobotWale, we grade robotics claims by shipping hardware first, pilot deployments second, and public announcements last. Foundation models accelerate development cycles, but they do not bypass the physics of actuation, safety certification, or supply chain constraints that dictate real-world adoption.
What Defines a Robotics Foundation Model?
- Multimodal input fusion: synchronized camera feeds, lidar point clouds, joint torque/position, and tactile signals are normalized into a shared latent space.
- Policy output: continuous motor commands (joint velocities, forces, or impedance parameters) rather than discrete token generation.
- Generalization metric: measured by success rate across unseen tasks, not benchmark scores on static datasets.
- Training data pipeline: fleet telemetry, teleoperation logs, and simulation-to-reality transfer with domain randomization.
- Deployment stack: edge inference optimization, real-time control loops at 500Hz+, and fail-safe fallbacks for safety-critical operations.
Grading the Contenders by Deployment Stage
Three architectures dominate current discussions: Figure’s Pi, Google DeepMind’s RT-2, and NVIDIA’s Groot. Each occupies a different point on the maturity curve when evaluated against shipping hardware, pilot deployments, and announcements.
Figure’s Pi: From Announcement to Shipping Hardware
Figure AI’s Pi models (Pi-0 and Pi-1) are trained on data collected from Figure 01 units operating in industrial settings. The company has moved beyond simulation-only training by logging real-world teleoperation and autonomy data across partner facilities. Figure has shipped Figure 01 hardware to multiple pilot sites, including manufacturing and logistics environments, where the Pi policy handles object manipulation, navigation, and task sequencing without hard-coded scripts.
Grading against our hierarchy: shipping hardware takes precedence. Figure 01 units are physically deployed, and Pi is actively running on edge compute in those units. Pilot deployments are extensive, with documented task success rates improving as fleet data accumulates. Public announcements remain forward-looking, but the hardware-first approach grounds their claims. The model is not a standalone product; it is licensed as part of the Figure platform, which bundles actuators, sensors, compute, and safety systems.
Google DeepMind’s RT-2: Research Milestones and Pilot Constraints
RT-2 introduced a VLA architecture that treats robot actions as language tokens, enabling cross-task generalization through pretrained vision-language models. Google released open weights and demonstrated manipulation capabilities on simulated and lab-scale platforms. The research is technically sound, and the paper provides reproducible training pipelines. However, RT-2 has not shipped as integrated hardware. Google’s deployments remain in research labs and academic pilot programs, with limited commercial integration.
Grading: announcements and research papers lead the maturity curve. Pilot deployments exist but are confined to controlled environments and university partnerships. Shipping hardware is absent. RT-2’s value lies in its architectural influence and open weights, which have accelerated industry-wide VLA research. Commercial viability depends on whether partners can adapt RT-2 to industrial-grade safety, latency, and certification requirements.
NVIDIA’s Groot and the Simulation-to-Reality Pipeline
NVIDIA’s Groot (evolved from Isaac Groot) focuses on large-scale simulation training, domain randomization, and real-world deployment via Isaac ROS and Isaac Sim. The platform emphasizes data generation at scale, policy training in physics-accurate environments, and edge inference on Jetson hardware. NVIDIA has partnered with integrators to run pilot programs in warehouses and factories, but Groot is fundamentally a software stack, not a robot.
Grading: pilot deployments and simulation pipelines lead. NVIDIA ships Jetson modules and Isaac licenses, not complete humanoid platforms. Real-world deployment depends on third-party robot manufacturers integrating Isaac ROS and Groot policies. The strength lies in reproducibility and compute scalability; the weakness is the dependency on hardware partners for physical deployment and safety certification.
India Availability, Hardware Integration, and Pricing
Foundation models do not ship as standalone products. They are integrated into robot platforms, licensed via API or enterprise agreements, or deployed on-premises using edge compute. India’s robotics market follows this pattern, with hardware imports, local integration, and software licensing driving costs.
Software Licensing vs. Landed Robot Costs
- Foundation model licensing: typically structured as annual enterprise subscriptions or revenue-sharing agreements. Estimated cost: INR 15 lakh to 50 lakh per year, depending on fleet size and inference tier. (Landed cost estimate, flagged as approximate and subject to vendor negotiation.)
- Edge compute modules: Jetson Orin or equivalent industrial GPUs. Estimated cost: INR 2.5 lakh to 4 lakh per unit. (Landed cost estimate, flagged as approximate.)
- Humanoid robot platforms: Figure 01, Agility Digit, or Boston Dynamics Spot (non-humanoid but relevant for pilot comparison). Landed cost in India ranges from INR 80 lakh to 2.5 crore per unit, including customs, GST, and integration. (Landed cost estimate, flagged as approximate.)
- Pilot program costs: facility adaptation, safety audits, and operator training typically add INR 10 lakh to 30 lakh upfront. (Landed cost estimate, flagged as approximate.)
Where to Buy or Pilot in India
Indian enterprises access foundation models through three pathways:
- Direct vendor partnerships: Figure AI, NVIDIA, and DeepMind affiliates offer pilot programs. Contact is routed through regional sales and enterprise agreements.
- Local system integrators: Indian robotics firms handle hardware import, safety certification, and on-site deployment. They bundle model licenses with actuator servicing and control software.
- Cloud inference tiers: Some vendors offer remote policy inference for early-stage pilots. Latency and data residency requirements limit this to non-critical tasks.
Procurement follows standard industrial equipment import rules. GST applies at 18% on hardware and software licenses. Data localization for AI models may require on-prem inference or approved cloud regions. Enterprises should request vendor spec sheets, pilot success metrics, and safety compliance documentation before committing.
The Path to a General Policy
A general policy in robotics requires three conditions: robust safety certification, consistent actuation performance, and continuous fleet learning. Foundation models accelerate the first two by reducing manual programming, but they do not eliminate the need for mechanical reliability and regulatory approval. Figure leads in shipping hardware with active pilots. NVIDIA leads in simulation and compute scalability. Google leads in open research and architectural influence.
The race is no longer about which model can manipulate a novel object in a lab. It is about which architecture can maintain sub-10ms control loops, pass industrial safety standards, and reduce total cost of ownership in Indian manufacturing and logistics environments. Hardware shipments will dictate the next phase. Pilots will validate generalization. Announcements will follow.
References
- Figure AI. Pi-0: Foundation Model for Robot Manipulation. https://www.figure.ai/blog/pi-0
- Google DeepMind. RT-2: A General Policy for Robotics. https://deepmind.google/discover/blog/rt-2-a-general-policy-for-robotics/
- NVIDIA. Isaac Groot: Foundation Models for Robot Learning. https://developer.nvidia.com/isaac/groot
- IEEE Spectrum. Robotics Foundation Models: Progress and Pitfalls. https://spectrum.ieee.org/robotics-foundation-models
- Boston Dynamics. Spot Platform Specifications and Pricing. https://www.bostondynamics.com/spot
- Agility Robotics. Digit Platform Technical Overview. https://www.agilityrobotics.com/digit


