The Latest Breakthroughs in Robotics: September 2026
A robot can perform individual tasks – like learn to fold a towel and more particular objects. But what happens when the towel is different, the table is lower, or someone moves it halfway through the task?
That comes to teaching them how to handle what they haven’t seen before.
The latest developments show this direction at several levels at once: robots are getting better at transferring skills to new environments, learning from demonstrations, dealing with physical contact, and adapting their behavior on the fly. At the same time, safety, easier programming, and hardware built for real-world work are becoming just as important as intelligence itself. If those mechanisms are unfamiliar, start with how robots move and learn: SLAM, whole-body control, reinforcement learning and behavioral cloning.
Developments we are going to look at are all about making robots more adaptable once they leave the lab.
TL;DR: Robots are getting better at handling what they haven’t seen before. Figure is testing household skills across unfamiliar homes; Skild and Generalist let robots pick up tasks from short demonstrations; Facet-0 and Safe-Stop tackle physical interaction and safety. Meanwhile, Agility is upgrading Digit for industrial work, FANUC and Google are automating robot programming, and OpenAI is moving toward building robots of its own.
Robot Learning & Foundation Models
Figure’s Helix 2.5 brings household skills to 30 new homes
Figure tested Helix 2.5 on tidying toys, folding towels, and making beds in 30 homes. Pretraining on Index, its dataset of human behavior, raised complete-task success from 9% to 56%. The robot learned the tasks elsewhere, then tackled the new homes and objects without additional fine-tuning. Helix 2.5 is a vision-language-action model, the same class as π0 and SmolVLA. This gives a concrete measure of how well household skills transfer beyond the places where robots learn them. Figure’s blog post →
Skild S1 learns from a single video prompt
Show S1 a demonstration, and it uses the video in its context window to guide the robot without updating its weights. Among examples there are potting plants, making pancakes, and brewing coffee, with tasks lasting up to ten minutes. On unseen tasks, Skild reports 66% versus 9% for a comparable language-prompted model, both trained on 100,000 hours of data. That metric measures cumulative per-step success, with human interventions used to recover from failures, mainly for the language-prompted baseline; it should not be read as a 66% fully autonomous completion rate. Skild AI blog post →
Facet-0 helps robots spot when a part gets stuck
Assembling a computer requires more than seeing where a component belongs. A robot also needs to handle the forces that appear when parts touch. Facet-0 combines images, instructions, and wrist force-and-torque measurements to predict both movements and their contact effects. It then improves through real-world attempts. Across five precision computer-assembly tasks, the full, task-adapted system achieved 82% average success, compared with 15% for the strongest tested baseline. These are research results from a controlled setup. Read more →
Safe-Stop teaches humanoids to assess whether they can stop without falling
An emergency stop is complicated when a robot is running, turning, or already losing balance. Safe-Stop estimates whether the robot can still reach a stable standing position. When both estimators indicate that stopping is feasible, it uses a learned stopping controller; otherwise, it switches to a damping fallback for the fall. In Unitree G1 simulations, the stopping policy achieved 96.4% success across 179,650 valid episodes starting from unfamiliar motion-capture states. The work tackles a practical question: what should a moving humanoid actually do when someone presses stop? Read more →
Generalist GEN-1.5 picks up short tasks from seconds of demos
GEN-1.5 can use 3–12 seconds of demonstration in its context to attempt tasks such as opening a zipper or unscrewing a jar. Generalist reports 59% average success across ten tasks without weight updates, rising to 83% after ten fine-tuning steps using five minutes of data per task. The prompts generally contain sensor data and action trajectories, rather than just ordinary video. These remain short, relatively simple tasks, but the amount of task-specific training is small. Read more →
Humanoid Robots in Industry 2026
Agility’s Digit 5 lifts 22.7 kg, recharges itself in 9 minutes, and introduces a new safety system and new hardware
Digit 5 is designed to repeatedly lift 22.7 kg, run for 90 minutes, and recharge in nine minutes. Swappable grippers expand the range of jobs it can tackle. Its new safety architecture detects nearby people and uses an independent controller to trigger avoidance, stopping, or sitting responses. Agility was also the first humanoid partner for NVIDIA's Halos safety stack, announced in June 2026. Agility says the design draws on more than 65,000 operating hours from Digit 4. Early access is expected in the first half of 2027, with general availability by year-end. Agility Robotics →
FANUC and Google turn engineering drawings into robot welding programs
FANUC’s new AI Welding Agent, developed with Google using Gemini Enterprise, reads a component drawing and generates welding settings, including current and voltage, along with the robot’s motion program. Operators capture the drawing using the camera already built into the CRX robot’s tablet controller, then can review or adjust the generated instructions before welding. The aim is to reduce the manual programming needed for each new part. FANUC scheduled demonstrations for September’s International Welding Show and shipments for the end of December 2026. FANUC →
Industry Moves
OpenAI is planning to build its own humanoid robot
Sam Altman confirmed on the Sources podcast that OpenAI plans to build humanoids and robots with other body designs. Infrastructure and manufacturing come first; personal robots for the home are a longer-term ambition. The company is also expanding its hardware team: Forbes reported 19 robotics openings on OpenAI’s careers page, covering areas such as actuators, firmware, and prototyping. No launch date, unit target, or manufacturing partner has been announced. Read more →
Sources and further reading
If you're just getting started with ML and AI, check out our curated list of Top 10 GitHub repos for AI & ML practitioners— collections of courses, guides, and projects to build your foundations.
From Turing Post
FAQ
How are robots learning new tasks in 2026?
Several approaches are emerging. Robots can learn from large-scale human video during pretraining, use demonstrations directly in their context, adapt from a few minutes of task-specific data, or combine visual information with physical signals such as force and torque. Skild’s S1, for example, uses a single video demonstration as a prompt for tasks it did not encounter during training, without updating its weights.
What are robot foundation models?
Robot foundation models are models trained across broad sets of tasks, environments, or embodiments so that their knowledge can transfer to new robotic tasks. They can connect vision, language, video, sensor information, and actions, reducing the need to build a separate model or policy from scratch for every task.
How did you like it?
Thank you for reading! [Sign-off — e.g. “Share this list with a friend who is still deploying models by hand.”]