Xiaomi-Robotics-1

Original Article ↗ Hacker News Discussion ↗

Xiaomi’s new Xiaomi-Robotics-1 project showcases an open, relatively small (≈10B parameters) vision-language-action model controlling robots that sort, fold, and pack laundry, impressing many with uncut footage of complex bimanual manipulation of deformable objects. Commenters see it as an important step toward practical household robotics and a potential analogue to how LLMs transformed knowledge work, while others stress that today’s systems are still slow, brittle demos far from affordable, fully autonomous home helpers. The thread branches into broader questions about job displacement, human reliance on automation, China’s growing role in open AI/robotics, and whether humanoid general-purpose robots will beat specialized machines and redesigned environments.

Technical approach and capabilities

  • Xiaomi-Robotics-1 is described as a relatively small (~10B parameter) vision–language–action (VLA) model yet capable of nontrivial manipulation.
  • It follows a “full VLA” end-to-end paradigm with an implicit world model, rather than explicit state representations and trajectory rollouts.
  • Training includes large-scale non-robot data plus robot demos (often humans tele-operating standardized UMI grippers and cameras), aiming for “embodiment-free” gripper/camera usage.
  • Reported benchmarks: improves Robodojo success (e.g., 14% vs prior 9%); transfer learning claims ~75% success on new complex tasks from <10 hours of demos, though memory tasks are weaker and results are statistically thin.

Generalization and hardware concerns

  • Discussion on how well such models transfer to other robots: likely decent for similar platforms (1–2 arms, 2-finger hands) but fine-tuning per arm type is expected.
  • “Embodiment-free” here mostly means standardized end-effector/camera, not truly arbitrary robot bodies.

Demo reliability and performance

  • Some distrust the main promo video due to frequent cuts; others point out an “uncut” 2×-speed full run that shows occasional loops and hesitations.
  • Several commenters warn that “uncut” still means “best run,” not average performance.
  • From the paper, one commenter notes roughly ~50% success on some tasks; others emphasize how hard deformable objects, thin affordances (zippers), and bimanual coordination are.

Laundry folding as benchmark

  • Many see laundry folding as a strong testbed: deformable, entangled items, occlusions, precision grasps, and multi-step sequencing.
  • Some argue it’s chosen because it’s visually intuitive and shows dexterity; others say it’s a very minor real-world time sink, especially without kids.

Household automation value vs cost

  • Enthusiasts are excited about offloading hated chores (laundry, dishes, tidying) and imagine future integration with Home Assistant, robot vacuums, and home inventory systems.
  • Skeptics note that current robots are likely more expensive and less capable than human cleaners for years, and that these are still controlled demos in constrained environments.

Broader AI and societal implications

  • Optimistic voices compare this to early cars: slow and clumsy now but on an exponential trajectory, with parallels to how LLMs rapidly improved.
  • Pessimistic voices foresee job displacement across low- and high-income work, concentration of robot ownership, and possible social decay (Wall‑E–style passivity).
  • There is meta-debate between “AI enthusiasts” who feel super-empowered and critics who see this attitude as manic or detached from societal risks.

China, open source, and geopolitics

  • Many highlight that recent Chinese open-weight models and robotics work are a major global positive, pushing openness in a space where US firms tend to be closed.
  • Others stress this openness is strategic, not guaranteed long-term; they call for more open-source effort from Western actors to avoid dependence.
  • The thread includes side arguments over human rights in China vs the West and accusations of both anti-China sentiment and pro-China bias on HN.

Robot form factor debate

  • Some argue non-humanoid, task-specific robots (snakes, spiders, swarms, dish-cube appliances) are more efficient than humanoids.
  • Others counter that humanoid-ish forms are increasingly viable and can exploit the huge installed base of human-optimized environments and tools, making redesigning the world for special-purpose bots less realistic.