Humanoid Robot Imitation Learning

Capital is flowing to the data layer beneath humanoid imitation learning

The same $60M Series B shows up across Axios Pro Rata and StrictlyVC with Sequoia, Nvidia and M12 (plus Microsoft, Qualcomm and Samsung in one report), and Sequoia's $60M also appears in a Parsers VC item that cites Mecka as human-motion data infrastructure. Mecka's pitch is essential data infrastructure for robotics, which suggests investors are paying for scalable demonstration and human-motion data rather than only for robot policies. AgiBot World, the largest source in the RoboJEPA mixture at 167.5k episodes and 2,734 action hours, reinforces that demonstration data is the scarce asset. Capital for the 28 days is $180M across 5 deals, with Sequoia in 3 and Nvidia and a16z each in 2.

Physical Intelligence's π0/π0.5 is the reference policy in nearly every paper in the feed, used as the real-world base on DROID and as the comparison target. Papers keep documenting its brittleness: 0.0% on LIBERO-Pro, 0% grasping on a moving conveyor beyond 3 cm/s, and 4/5 dropped coats with higher latency than GR00T. Challengers such as Long-WAM, Recova and VioLA are framed explicitly against it.

Why it matters · Being the incumbent baseline brings mindshare but also scrutiny, and displacing π0.5 on robustness is the clearest opening for new entrants.

One paper argues discrete action parametrizations match or beat continuous ones across imitation learning, multi-task VLA training and real-robot tasks, directly challenging π0, π0.5 and GR00T N1. Another (H-VLA) argues direct language-vision-to-action mapping has a generalization ceiling. SkipVLA's hybrid of a VLA with a classical planner cuts latency by 59.6% on Jetson Thor. The field is questioning architecture fundamentals rather than just scaling.

Why it matters · Architecture choices that looked settled are open again, which lowers lock-in risk for new humanoid-focused startups.

Long-WAM reaches 90-100% grasping success on the Unitree G1 humanoid, while VioLA beats GR00T N1.7, which scores 16.7% on locomotion and 0/35 on manipulation. Carnegie Mellon's Sreenath and Shi labs, working on bipedal control and robot learning, anchor the academic side. Mondo Robotics is co-developing foundation world action models for whole-body control. World-action models are becoming the competing path to pure VLA scaling.

Why it matters · Open humanoid baselines are weak, so well-executed whole-body methods can win visible benchmark leads quickly.

EgoLAP trains robots from egocentric human video using language-action reasoning, following the π0.5 architecture, and mixes in 25% bimanual robot data from AgiBot and others. This blends cheap human data with scarcer robot demonstrations. Mecka's human-motion data focus fits the same direction.

Why it matters · Cheaper human-motion data lowers the cost per demonstration and favors firms that can capture it at scale.