Humanoid Imitation Sim-to-Real Fusion
Companies and labs that simultaneously use human demonstration data (imitation learning) and large-scale simulation with sim-to-real transfer to train humanoid robots, creating a tightly coupled pipeline that neither parent theme alone captures.
CAPITAL FIGURES ARE MEDIA-EXTRACTED ESTIMATES, NOT VERIFIED FILINGS.
EXTRACTED FROM 25+ PODCASTS & VC NEWSLETTERS · MEDIA-REPORTED FIGURES, NOT VERIFIED FILINGS
World action models fuse demonstration data with simulated dynamics
Shanghai AI Laboratory's InternW0 (Sept 23) and InternW0-Δ (Sept 25) decouple a video prediction expert from an action expert, which is the closest published template of a coupled imitation and world-simulation pipeline. InternW0-Δ is pretrained on a 20K-hour corpus built from real demonstrations (AgiBotWorld, Galaxea's 500+ hours, ActionNet, RealSource, Fourier). It reaches 30 Hz control with 152.8 ms latency on a single RTX 5090. The lab is open-sourcing the full stack, including code, weights, and data pipeline. The result is that imitation data and learned simulation are trained together rather than as separate tracks.
Figure's Helix, a hierarchical VLA, is repeatedly used as the benchmark in the research literature. Papers cite its dual-frequency design and Helix 02's two-level whole-body loco-manipulation, and one paper critiques its fixed-timer termination. Helix 2.5 (Sept 2026) was framed as passing the Woz test with zero-shot whole-body autonomy in an unfamiliar home. F.03 arrived at BMW, so the stack is already in production. Practitioners also point to strong low-data performance from in-context learning at Figure.
Why it matters · Figure's architecture sets the design target, and its published weaknesses mark where competitors can differentiate.
Weekly capital is lumpy: $117M across 4 deals in the week of Aug 24, nothing across the next several weeks, then $72M across 2 deals in the week of Oct 12. The latest round is a $36M Series B backed by Ascensia, reported by both StrictlyVC and PitchBook on Oct 13. The stage mix over 90 days shows Series A and B deals, plus a $5M Contrary seed and a $20M Lachy Groom Series A, rather than only late-stage mega-rounds. Ascensia is the top investor in the theme with 2 deal counts.
Why it matters · A cluster of seed to Series B rounds suggests investors are still underwriting new entrants in this fusion niche, though the data is too thin to confirm concentration.
1X's humanoid world-model dataset (23.5k episodes, 90 hours) is part of the RoboJEPA training mixture, which shows company data flowing into shared multi-source corpora. This matches InternW0-Δ's pooling of AgiBot, Fourier, Galaxea, and RealSource data. Practitioners on Lightcone call cross-embodiment generalization deeply underappreciated.
Why it matters · Whoever aggregates the most diverse embodiment data can train a transferable policy, which favors data-sharing ecosystems over closed single-robot fleets.
A recent arXiv paper argues that Figure, Boston Dynamics, and Apptronik over-focus on end-effector dexterity, and that using the whole body as a contact surface is the bigger unlock. A related paper (WT-UMI) flags bulky-object handling as a core use case for Figure's logistics and manufacturing targets. A My First Million discussion argues that intelligence, not manufacturing, is the real bottleneck for humanoids.
Why it matters · Demonstration collection methods and simulation contact models will have to cover the whole body to unlock logistics deployments.