What changed
At Long Horizon in San Francisco, Felipe Chavez Cortes, co-founder and CEO of Robot.com, discussed the challenges of deploying robots, highlighting the need for AI integration to reduce manual work. Recent advancements in AI systems, such as VoLo and POT-VLA, show progress in automating tasks like action completion, object tracking, and error correction.
What this means in practice
VoLo combines π0.5 with a vision-language model to plan and monitor execution, improving task success from 6 of 42 trials to 18 of 42 across 14 tasks. POT-VLA maintains persistent 3D records of objects to guide actions and check relations, while BATON records constraints on skills and their transitions to ensure coordinated actions. Generalist's GEN-1.5 accepts physical demonstrations as prompts, achieving 59% average success with one 3–12 second demonstration and 83% with five minutes of data and ten fine-tuning steps. Skild's S1 uses egocentric video to communicate task organization, leveraging pre-existing skills.
Limitations and availability
Despite these improvements, success rates remain below 50% in some cases, and the systems still require human oversight for diagnosing failures. For example, a successful grasp may fail due to physical constraints, requiring a different approach. Research like Recursive Harness Distillation and Google's Gemini 4 Argon robotics demonstrations show AI systems beginning to handle some diagnostic tasks, but full autonomy remains a challenge. The supplied announcement does not specify exact release dates or availability for commercial use.
