Eka Robotics raises $150m for vision-force-action model

Eka Robotics, a startup founded by MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja, has closed a $150 million Series A funding round. The investment, led by a16z and F-Prime, values the company at $1.45 billion post-money and highlights a significant shift in how artificial intelligence interacts with the physical world through advanced manipulation architectures.

What happened

The robotics sector has long struggled with the simple act of picking up objects, despite massive investments in deep learning models trained on video data. Eka Robotics claims to have solved this bottleneck by developing a system that exceeds human performance in specific dexterous tasks. According to Raghu Raghuram, an a16z general partner, the technology is already demonstrating superhuman capabilities and is expected to expand to many more manipulation tasks in the coming months.

Industry experts have reacted with surprise to the demonstrations. Kris Hauser, a professor at the University of Illinois Urbana-Champaign, noted the system's ability to adapt to diverse objects without retraining. He described a distinct "humanness" in the robot's behavior, particularly its persistence when initial attempts fail. Another observer from a frontier laboratory compared the emotional impact of seeing Eka’s robot to the breakthrough moments in legged locomotion seen five to ten years ago, citing the smoothness and speed of its movements as particularly striking.

Performance metrics provided by the company indicate that the robot can pick up approximately 1,500 objects weighing up to 1 kg in an hour, significantly outpacing the human average of 600. The system reportedly achieves a reliability rate of over 95 percent. The speed and force of the arm are substantial enough that the algorithm must possess high confidence in its control logic to avoid self-damage, a level of trust that suggests extensive testing and robust underlying models.

How it works

The core innovation lies in shifting from vision-language-action models to a vision-force-action architecture. While traditional models rely heavily on visual inputs to determine action, Eka’s approach prioritizes force feedback. Pulkit Agrawal explains that vision only provides rough location and pose data, whereas the actual grasping and searching processes are driven by force interactions. This aligns with critiques from experts like Rodney Brooks, who argue that generative AI robotics fails without representing touch and force.

To enable this, Eka designed a custom robotic hand with greater flexibility and a higher density of force sensors than traditional grippers. This hardware allows the robot to react instantly to contact, preventing damage to fragile items like raspberries. The training process leverages large-scale physics simulations to model the multiple forces encountered during grasping. By using reinforcement learning within these simulated environments, the team can generate thousands of hours of robot experience in minutes, bypassing the limitations of collecting real-world demonstration data.

This bottom-up approach contrasts with the top-down planning methods used by many general-purpose AI models. While systems like Astra may handle high-level decision-making, they often lack the physical precision required for delicate tasks. Eka’s model handles the physical execution, allowing it to bridge the gap between general intelligence and reliable physical performance. The simulation-based training enables the robot to surpass the capabilities demonstrated by human operators, as it is not constrained by the limited data available from manual demonstrations.

Key details

• Eka Robotics raised $150 million in Series A funding, reaching a $1.45 billion post-money valuation.
• The founding team includes MIT professor Pulkit Agrawal and former Google DeepMind researcher Tuomas Haarnoja.
• The robot utilizes a vision-force-action model rather than the more common vision-language-action framework.
• Performance tests show the robot can pick up 1,500 objects per hour with over 95 percent reliability.
• Training relies on scalable physics simulations and reinforcement learning instead of human demonstration videos.
• Current industrial pilots focus on pick-and-place operations and wire manipulation in secretive deployments.

Why it matters

For software engineers and technical leads, this development signals a move away from purely data-driven visual models toward hybrid systems that integrate physical laws into the learning loop. The reliance on simulation for reinforcement learning reduces the dependency on expensive and slow real-world data collection. This shift could accelerate the deployment of robots in unstructured environments where visual data alone is insufficient for precise manipulation.

The distinction between high-level planning and low-level physical control is becoming clearer. General AI models may dictate what needs to be done, but specialized physical AI models like Eka’s will execute the task safely and efficiently. This separation of concerns allows for more robust systems where the physical layer can operate with high speed and confidence, trusting the algorithm to prevent hardware damage while interacting with delicate or variable objects.

However, challenges remain in transitioning from controlled lab environments to industrial assembly lines. Reliability over extended periods, such as running without failure for a year, is a significant hurdle. Additionally, the current focus on pick-and-place and wire handling must expand to more complex value-added tasks to prove broader applicability. The durability of the hardware, currently rated for 5 million duty cycles, also needs to meet the 20 million cycle standard of industrial arms.

What you can do

• Evaluate whether your current robotics projects rely too heavily on visual data without adequate force feedback integration.
• Explore reinforcement learning frameworks that support physics-based simulations for training manipulation policies.
• Investigate sensor-rich gripper designs that can provide high-resolution force data for closed-loop control.
• Consider hybrid architectures where high-level LLMs handle task planning while specialized models manage physical execution.
• Monitor developments in simulation-to-real transfer techniques to reduce the need for extensive real-world testing.
• Assess the durability requirements of your hardware against the operational demands of continuous industrial use.