Locomotion

Unitree G1 is the default humanoid research platform

Across the last six weeks of signals, the Unitree G1 is the hardware used for real-world validation in diffusion-policy generation, human-to-humanoid retargeting, depth-based locomotion, VLM orchestration (GPT-6 Astra) and conveyor grasping. Unitree also ships the UnifoLM-WBT dataset (4.45M frames), which makes it both the hardware and a data standard. A podcast guest (Lindon Gao of Dyna) credits Unitree with shrinking the sim-to-real gap. Separately, a Chinese podcast cites a roughly 200B RMB valuation for Unitree against 20x growth math, so the platform's dominance is being weighed against how much value is left to capture.

Multiple papers deploy policies on a real G1 with no fine-tuning, including a weight-space diffusion model that reaches best-of-100 success of 1.000 on walking, dancing and spin kick, and a unified depth-based policy running at 50 Hz. Training relies on GPU-parallel simulation (4,096 environments, MuJoCo-based mjlab and MuJoCo Warp), and one framework transfers to the G1 without any G1-specific training.

Why it matters · As transfer gets cheaper, the differentiator moves from locomotion controllers to simulation compute, datasets and the tasks built on top.

A paper argues that Figure AI, Boston Dynamics and Apptronik focus on end-effector dexterity, while the bigger unlock is using the whole body as a contact surface. Supporting work includes RobotMover moving heavy tables and racks, ANYmal door traversal, and Spot learning manipulation from human motion capture. A single policy handles 13 terrains in ANYbotics experiments.

Why it matters · Winners may be the teams that unify legs, torso and arms in one policy, not those with the best hands.

ResSafe separates a task-optimized policy from a residual safety policy. InterEvolve evolves reward programs at test time using an LLM agent plus CMA-ES, so the humanoid handles unseen tasks without retraining the controller. Off-policy AWR fine-tuning is reported as far more sample-efficient than PPO.

Why it matters · Safety and adaptation tooling is a potential standalone product category sitting above the locomotion stack.

Funding was concentrated in two weeks: $117M across 4 deals in the week of 08-24 and $72M across 2 deals in the week of 10-12, with zero in most other weeks. The October money is a $36M Series B led by Ascensia, reported by both StrictlyVC and PitchBook, so the two deal counts overlap. Commercial signals continue: Boston Dynamics' Stretch at DHL, Agility's broader Amazon relationship, and Shift Robotics' Moonwalkers consumer shoes.

Why it matters · Capital is lumpy and selective, so deployment traction is likely to decide who gets funded next.