Robotics
On October 7, 2025, Tesla unveiled Optimus Gen 3 performing Kung Fu, cooking, and cleaning—all learned through observation, not coding. Figure 02 completed a 20-hour continuous shift. Boston Dynamics began testing Atlas at Hyundai's Georgia facility. The investment tells the story: Figure AI raised $1 billion in 2025. Apptronik raised $403 million. Agility Robotics raised $400 million. 2025 became the breakthrough year for humanoid robots.
The Humanoid Race
Tesla Optimus
Tesla plans to produce 5,000-10,000 Optimus units in 2025, with parts procurement for up to 12,000. The robot stands 1.73m tall, weighs 57kg, and can carry 20kg while walking or lift 68kg. Target price: $20,000-$30,000—"less than a car."
Unlike research robots, Optimus is designed for mass manufacturing. Tesla applies its expertise in electric powertrains, AI development from Full Self-Driving, and production scale.
Boston Dynamics Atlas
Boston Dynamics repositioned Atlas in April 2024 as all-electric, retiring the hydraulic version. New Atlas: 1.5m tall, 89kg, about 28 degrees of freedom. Pilot testing at Hyundai's Georgia facility; commercial launch 2026-2028. Estimated price: $140,000-$150,000. Collaboration with Toyota Research Institute on AI-driven behavioral models.
Figure AI
Figure 02 blends large language models with motor control for natural language tasking. The 20-hour continuous shift demonstrated industrial endurance. Figure's Helix AI platform set to debut in homes in 2025.
Embodied AI: The Foundation Model Era
Robotics has entered the foundation model era. Instead of custom perception stacks and task-specific controllers, robots are increasingly powered by large multimodal models that interpret scenes, understand instructions, and produce structured actions.
Open X-Embodiment Dataset
The Open X-Embodiment dataset pools 60 robot datasets from 34 labs worldwide, demonstrating 500+ skills and 150,000+ tasks across 1 million+ episodes. Developed with academic labs across 20+ institutions.
RT-1-X showed 50% success rate improvement on average across five commonly used robots. Training RT-2 on multiple embodiments tripled its performance on real-world robotic skills. World-model pretraining with optic-flow action representations enables >50% policy improvement with minimal new-target data.
Expanding the number of training embodiments yields more effective generalization than increasing trajectory count for a fixed embodiment set.
GEN-0: Scaling Physical Interaction
GEN-0 is pretrained on over 270,000 hours of real-world manipulation data, growing at 10,000 hours per week. Works across different robots by design—tested on 6DoF, 7DoF, and 16+ DoF semi-humanoid systems.
This marks "the beginning of a new era: embodied foundation models whose capabilities predictably scale with physical interaction data—not just from text, images, or simulation—but the real world."
Vision-Language-Action Models
VLAMs have become the clearest expression of how foundation models infuse new life into robotics. End-to-end approaches like GR-3 and Gemini Robotics enable models to internalize geometry, affordances, and contact dynamics through joint training of perception, semantics, and control.
The ROS Ecosystem
The Robot Operating System (ROS) remains the backbone of robotics research. ROS 2 addressed real-time requirements, security, and multi-robot coordination for production deployments.
2025 Developments
- RSS 2025 Best Paper - CMU's DexWild, a ROS 2 package for dexterous human interaction policies
- Medical Robotics - Interoperability between 3D Slicer, ROS2, Gazebo, and da Vinci Research Kit at ISMR 2025
- DexHand - Low-cost, open source dexterous hand with ROS 2 drivers and ChatGPT integration demos
- Unitree Open Source - Imitation learning (Diffusion Policy, ACT) for G1 humanoid, Z1 arm, and Dex3 hand
Where to Find Robotics Research
Preprints and Papers
- cs.RO on arXiv - Robotics preprints
- OpenReview - CoRL, RSS, ICRA submissions
Open Source
- ROS - Robot Operating System
- Open X-Embodiment - Cross-robot datasets and RT-X models
- LeRobot - Hugging Face's robot learning framework
Why Robotics Shares
Complexity. No one lab can build everything. Robots need perception, planning, control, simulation, and hardware. Open source lets labs specialize while sharing infrastructure.
Data hunger. Foundation models need massive datasets. The Open X-Embodiment dataset exists because 34 labs pooled their data. No single organization could collect 270,000 hours of manipulation data alone.
Standards. ROS became the de facto standard precisely because it was open. Proprietary alternatives couldn't achieve the same network effects.
The result: a field where the fundamental software infrastructure is free, where research papers include code, and where startups building $20,000 humanoids can leverage the same algorithms as billion-dollar research labs. The humanoid moment arrives on open source rails.