Humanoid Imitation-Sim Hybrid Learning
Companies and labs that jointly leverage imitation learning from human demonstrations AND sim-to-real transfer to train humanoid robots, using simulation to scale and diversify the teleoperation/motion-capture data that imitation learning alone cannot provide.
CAPITAL FIGURES ARE MEDIA-EXTRACTED ESTIMATES, NOT VERIFIED FILINGS.
EXTRACTED FROM 25+ PODCASTS & VC NEWSLETTERS · MEDIA-REPORTED FIGURES, NOT VERIFIED FILINGS
World-model pretraining on pooled demonstration data is the new humanoid recipe
Shanghai AI Laboratory's InternW0 and InternW0-Δ pair a video-prediction expert with an action expert, pretrained on a 20K-hour corpus that pools AgiBotWorld, RealSource, Fourier, Galaxea (500+ hours) and ActionNet data. This is the hybrid thesis in practice: real human and teleoperated demonstrations are the base, and learned dynamics models extend them beyond what imitation alone covers. The stack runs at 152.8 ms round-trip latency on a single RTX 5090, a 5.11x speedup that enables 30 Hz control. The lab is open-sourcing code, weights, and data pipelines, and 1X's 23.5k-episode world-model dataset also appears in the RoboJEPA training mixture.
Figure's Helix, Helix 02 and Helix 2.5 use a two-level, dual-frequency hierarchy for whole-body loco-manipulation. Helix 2.5 is described as zero-shot in an unfamiliar home, a claim that implies broad training diversity. Academic papers now critique Helix's fixed-timer termination, so the design is being scrutinized as well as copied.
Why it matters · A shared architecture makes training data and sim environments the main differentiator between humanoid players.
Practitioners call intelligence, not manufacturing, the limiting factor for humanoids, and cross-embodiment generalization is flagged as underappreciated. Figure is cited for strong low-data in-context learning. A paper argues that using the whole body as a contact surface is the unlock that most companies, including Apptronik and Figure, are missing. 1X is named as directly affected by WT-UMI-style data collection for deformable objects like laundry.
Why it matters · Demand is rising for tooling that turns scarce teleop and mocap data into diverse, transferable training sets.
Figure's $1B Series C at a $39B valuation and other growth rounds dominate headline dollars. Meanwhile, in the last 28 days a $36M Ascensia-led Series B was reported twice, along with a $20M Series A (Lachy Groom) and a $5M seed (Contrary). The weekly chart shows $117M across 4 deals in the week of 08-24 and $72M across 2 deals in the week of 10-12, with nothing in most other weeks.
Why it matters · Funding is lumpy and concentrated, so early-stage sim and data-infrastructure startups can get funded while platform leaders raise at far higher valuations.
Figure's F.03 is arriving at BMW, Apptronik has launched Apollo 2, and Google DeepMind has launched Gemini Robotics 2. 1X sells NEO as a $20,000 home robot, which turns real homes into a data source. Logistics sorting runs on learned models rather than traditional control.
Why it matters · Deployment creates real-world data that sim must then match, tightening the loop between imitation and sim-to-real.