Sim-to-Real Transfer

World models are being audited as policy simulators

Research is now testing whether learned world models can replace physical simulation for evaluating robot policies. DreamTrue's predicted success rates track ground-truth RoboTwin 2.0 simulation at Spearman rho 0.937, while the Ctrl-World baseline overestimates success with 26 false positives in 240 episodes. A reward model trained across eight world models (including GE-Sim 2.0, IRASim and Cosmos Predict 2.5) yielded a 44.9K-video corpus with 30.4K defect labels, in effect a comparative failure benchmark. 1X's world-model dataset (23.5k episodes, 90 hours) is already part of the RoboJEPA training mixture, so humanoid players supply the data these evaluators learn from.

Greylock backed Antioch, a simulation platform for physical AI, as a Series A in September. Applied Intuition's earlier Series A drew Lux, a16z and Elad Gil. Applied Intuition is described as a lighthouse-customer business with enormous ACV (GM is a customer), now expanding into defense, construction and mining. Capital in the past 90 days was concentrated in a few weeks, such as $117M in the week of Aug 24.

Why it matters · Simulation is being funded as picks-and-shovels infrastructure for physical AI, but the addressable market per vendor may be structurally small and concentrated.

ManiSkill3 cut training time from 7h54m on RLBench to 27 minutes with 32 parallel environments, an 18x speedup. Colosseum V2 is explicitly designed to supersede CoppeliaSim, which is limited to single-environment instances. RoboHorizon and ManiSkill3 launched together on Aug 4, signalling a throughput race.

Why it matters · Training throughput is becoming the competitive variable, and legacy simulators like CoppeliaSim risk losing relevance for large-scale RL.

Applied Intuition is cited as having invested heavily in neural reconstruction for sim-to-real transfer. The OmniDreams paper argues that paradigm is structurally limited for long-tail safety validation. The debate is shifting from visual fidelity to coverage of rare events in autonomous vehicle testing.

Why it matters · Incumbent simulation approaches may be vulnerable to generative alternatives, which creates openings for new entrants.

1X argues the humanoid form factor is a moat because it unlocks internet-scale human video for training. Research on whole-body contact and WT-UMI is flagged as directly relevant to 1X (deformable objects) and Apptronik (collaboration), whose Apollo 2 launched in August. Former 1X Chief Scientist Eric Jang independently converges on an 18-month timeline for household robots.

Why it matters · Real-world data advantages may outweigh pure simulation for humanoids, which shapes how much sim-to-real tooling vendors can capture.