Amazon FAR's FIBER trains a real Unitree G1 humanoid to chain loco-manipulation skills
CyberRobooo · x · 2026-10-24
FIBER, a project by Amazon FAR with CMU and Stanford, demonstrates versatile loco-manipulation on a real Unitree G1: catching randomly thrown boxes, crouching to carry boxes under low beams, picking boxes off boxes, and pushing open heavy doors, all from onboard depth sensing.
The key idea is planning in an action-aligned latent space: the planner outputs latent commands to a frozen behavior foundation model (BFM), where each latent is defined by the actions it applies. The BFM closes the loop at 50 Hz while the planner replans at 5 Hz. Experts, projectors, and the planner are trained entirely offline in simulation, then transferred to the real robot. Code coming soon.
More from Embodied
- Tesla starts production of Optimus humanoid robots, targets 20,000 units per week — LexiLove · 2026-10-24
- Amazon FAR, CMU and Stanford's FIBER trains Unitree G1 to chain skills for delivery runs — CyberRobooo · 2026-10-24
- iPhone Duo in stock for immediate pickup at French Apple Stores, from €3,089 — mitchdeg · 2026-10-24
- PWAs can't use foldable screens on the Duo yet, but two Chrome flags unlock the APIs — mitsuhiko · 2026-10-24
- Robot ducks trained in RTX Spark simulation flip live on stage at NVIDIA event — NandoDF · 2026-10-24
- Baidu's Apollo Go and Lyft's Freenow unveil RT6 robotaxi in London, targeting 2027 driverless launch — Baidu_Inc · 2026-10-24