Robotics Unicorn Sharpa and NVIDIA Aim to Bridge Physical AI’s Simulation Gap

(source: Sharpa)

AI robotics startup Sharpa announced new research developed with NVIDIA aimed at improving how robots are trained in simulation, one of the key bottlenecks in scaling physical AI.

The collaboration introduces Tacmap, a simulation framework designed to enable high-fidelity tactile simulation without sacrificing computational speed. Traditional robotics simulations often require a trade-off between realism and performance. Tacmap attempts to solve that problem through a shared geometric representation that allows both detailed physics and faster training cycles.

Sharpa plans to open-source the framework to broaden adoption across the robotics research community.

Simulation is central to Sharpa’s approach to developing robots capable of human-like dexterity. The company trains robots either through reinforcement learning in virtual environments or by generating synthetic data used to pre-train its Vision Tactile Language Model (VTLA). These models allow robots to learn manipulation tasks by combining tactile signals with visual and language-based understanding.

In parallel, NVIDIA’s GEAR Lab researchers demonstrated that policies derived from pre-training the GR00T model on more than 20,000 hours of human video could be transferred to robots equipped with Sharpa’s dexterous robotic hand, Wave. The system achieved a 54% higher success rate on manipulation tasks including assembling model cars, operating syringes, and sorting cards.

Sharpa’s hardware platform includes Wave, a human-scale robotic hand with 22 active degrees of freedom and tactile sensors, and North, a general-purpose humanoid robot designed for full-body manipulation tasks.

Founded in 2024 and headquartered in Singapore, Sharpa has quickly reached unicorn status and focuses on developing dexterity-first robotic systems designed for both enterprise and consumer environments. The company will showcase its technology at NVIDIA GTC 2026 in San Jose.

TheMarketAI Take

As we’ve discussed previously, physical AI faces a learning problem that software AI largely avoided.

Large language models can train on the internet. Robots cannot. Physical AI must learn from scarce real-world interactions involving vision, touch, motion, and physics. Collecting that data is slow and expensive.

Simulation is increasingly seen as the bridge.

High-quality virtual environments allow robots to run millions of training scenarios without the cost of real-world experimentation. The challenge has always been the “sim-to-real” gap: what works in simulation often fails when exposed to the unpredictability of the physical world.

Sharpa and NVIDIA’s work highlights the industry’s growing belief that better simulations may be the key to unlocking scalable robotics. If robots can reliably learn dexterity in virtual environments and transfer those skills into reality, the learning bottleneck that has slowed physical AI for decades may finally begin to narrow.

For now, the physics still win. But simulation may be how robotics finally catches up to software AI.