GEN-0: Embodied Foundation Models That Scale with Physical Interaction

ListenWhen Robots Ossify: GEN-0 and the Scaling Law for Physical Interaction10:43

Show notes

Generalist AI's GEN-0 is a family of embodied foundation models trained on 270,000 hours of real-world manipulation data — and it's the first robotics work to report *ossification*, where small models stop absorbing data past a certain scale. We dig into the ~7B phase transition, the clean pretraining-data power law on real-robot success, and Harmonic Reasoning as an alternative to System-1/System-2 stacks. Along the way we ask what a data-scaling bet from robotics means for how we think about video generation. **Sources** - GEN-0: Embodied Foundation Models That Scale with Physical Interaction (Generalist Team, blog)

Subscribe to the channel: https://podcast.lumalabs.link/27a25f51ea81d41a8253f46c6133a35e/feed.xml

TL;DR

GEN-0 is Generalist AI’s first-generation embodied foundation model family (up to 10B+ params) trained directly on 270,000 hours of real-world manipulation data collected across homes, warehouses, and workplaces, with the corpus growing at 10,000+ hours per week. Two headline empirical results sit above the rest: (1) a phase transition at ~7B parameters where smaller models exhibit LLM-style ossification in the high-data regime (weights fail to absorb new interaction data over training) while larger ones continue to improve — the first filed observation of ossification in robotics; (2) a strong pretraining-data power law on downstream post-training success, extending Chinchilla-style predictability to real-robot policies with peak per-task success ≈99%. The architecture introduces Harmonic Reasoning — asynchronous continuous-time interleaving of sensing and acting tokens — as an alternative to System-1/System-2 stacks and to inference-time real-time-chunking guidance for scaling thinking without stopping physics. Cross-embodiment coverage spans 6-DoF, 7-DoF, and 16+-DoF semi-humanoid platforms from one model.

Key claims

Ossification phase transition at ~7B parameters. At 1B, models “struggle to absorb complex and diverse sensorimotor data during pretraining — model weights become unable to absorb new information over time.” 6B begins to benefit from pretraining with strong multi-task behavior. 7B+ internalizes large-scale robotic pretraining data and transfers to downstream tasks with only “a few thousand steps of post-training.” Ossification had previously only been observed in LLMs at ~O(10M) scale; observing it in robotics at ~O(1B) scale is framed as consistent with Moravec’s paradox — physical commonsense has a higher compute activation threshold than abstract reasoning [§Surpassing the Intelligence Threshold].
Pretraining data → post-training success follows a predictable power law. Model checkpoints pretrained on varying subsets of the corpus, then post-trained with SFT jointly on 16 tasks, show downstream validation error following $L(D) = a \cdot D^{-b} + c$ (form given, exact fit constants not disclosed) — used to answer “how much pretraining data is needed to reach a target next-action prediction error” and to trade pretraining data against task-specific post-training data [§Scaling Laws for Robotics, Figs. 2, 4].
Real-robot success rates track pretraining scale under blind A/B evaluation. Increasing pretraining data monotonically lifts closed-loop success across tasks including clothes handling, with peak per-task success up to 99% when large-scale pretraining is combined with ample task-specific post-training data. Pretraining and post-training datasets were collected in different environments by different people — no overlap [§Scaling Laws for Robotics, Fig. 3].
Harmonic Reasoning trains an asynchronous, continuous-time interplay between streams of sensing tokens and acting tokens, allowing scaling to large model sizes without depending on System-1/System-2 architectures or on inference-time guidance methods (e.g. real-time chunking). Positioned as the answer to “physics doesn’t stop” — you cannot afford to pause the acting stream to think longer as in LLM inference-time scaling [§Introducing GEN-0].
Cross-embodiment by architecture. Same GEN-0 model has been tested on 6-DoF, 7-DoF, and 16+-DoF semi-humanoid robots — cross-embodiment is a design property, not a per-embodiment adapter [§Introducing GEN-0].
Data at 270,000 hours + 10,000 hours/week and accelerating. Corpus spans homes, bakeries, laundromats, warehouses, factories, powered by a global network of “1,000s of data collection devices and robots.” Custom hardware, dataloaders, dedicated Internet lines, multi-cloud contracts, and O(10K) CPU cores for multimodal processing let training absorb “6.85 years of real-world manipulation experience per day of training” [§Robotics is No Longer Limited By Data].
Data quality and diversity dominate volume; different pretraining mixtures yield systematically different downstream characteristics. Table 1 reports 8 mixtures × 10 long-horizon tasks × 3 dimensions (dexterity, applications, generalization) under two metrics: prediction MSE ( $||\mathbf{a}^\star - \hat{\mathbf{a}}||_2^2$ ) and mode-seeking reverse KL estimated via a Monte-Carlo Gaussian-mixture estimator over $M$ policy samples. Empirical read: models with low prediction error + low reverse KL transfer better under SFT post-training; models with high prediction error + low reverse KL are more distributionally multimodal and help post-training RL [§Science of Pretraining, Table 1].
Positioning of pretraining objectives. GEN-0 is framed as the first class of embodied foundation models whose capabilities predictably scale with real physical-interaction data — not text, not images, not simulation — with the implied contrast to synthetic / text-derived / sim-derived embodied-pretraining recipes.

Method

GEN-0 is a family of embodied foundation models (sizes explicitly named: 1B, 6B, 7B, “10B+”) built for multimodal training on high-fidelity raw physical interaction. Architecture inherits standard vision + language strengths, with Harmonic Reasoning as the distinguishing structural piece: an asynchronous, continuous-time coupling of sensing-token and acting-token streams that “trains the model to simultaneously think and act seamlessly” rather than alternating between them. The pretraining corpus is 270,000 hours of real-world manipulation captured across 1,000s of homes, warehouses, and workplaces via a globally distributed collection stack; corpus grows at ≥10,000 hours/week. Scaling experiments sweep model size at a fixed pretraining regime to identify the ossification / intelligence-threshold curve. Scaling-law experiments hold a fixed downstream post-training budget and vary the size $D$ of the pretraining subset, then fit $L(D) = a\,D^{-b} + c$ on downstream post-training loss. Real-robot scaling curves use blind A/B evaluations across sites with disjoint pretraining/post-training environments. Data-mixture ablations (Table 1) compare 8 mixtures sourced from different collection operations across dexterity / applications / generalization axes, using reverse KL (mode-seeking) alongside prediction MSE. Cross-embodiment evaluation is over 6-DoF / 7-DoF / 16+-DoF semi-humanoid platforms with the same base model.

Results

Ossification / phase transition: 1B fails to absorb the corpus (ossifies), 6B benefits with strong multi-task capabilities, 7B+ absorbs at scale and adapts to new tasks with “only a few thousand steps of post-training”; 10B+ shows continued improvement [§Surpassing the Intelligence Threshold].
Pretraining scale scaling law: monotone gains across all 16 SFT tasks as pretraining data increases; validation error fits a clean power law in $D$ (form provided; exponents not disclosed) [§Scaling Laws, Fig. 2, Fig. 4].
Peak real-robot performance: “up to 99% peak performance in certain cases” on real-robot A/B tests when large-scale pretraining is paired with ample per-task post-training data; monotone gains at fixed small (5.6-hour) post-training budgets as pretraining data grows [§Scaling Laws, Fig. 3].
Cross-embodiment: same model executes on 6-DoF, 7-DoF, and 16+-DoF semi-humanoid platforms [§Introducing GEN-0].
Data-mixture ablation (Table 1): 8 partner/class mixtures × 3 evaluation axes (Dexterity / Applications / Generalization) reported for both MSE and reverse-KL; ranking flips across metrics, empirically supporting “different mixtures produce different characteristic pretrained models” (SFT-friendly vs. RL-friendly).
Infrastructure: “6.85 years of real-world manipulation experience per day of training” absorbed via custom hardware + dataloaders + laid-fiber uplinks + O(10K)-core multimodal processing; dozens of PB compressed.
No head-to-head tables against π*0.6, GR00T, XR-1, Spirit-v1.5, Embodied-R1.5, LingBot-VA on shared benchmarks (LIBERO / RoboArena / ArmnetBench). Post is a research blog format.

Why it’s interesting

Direct predecessor to GEN-1.5: Embodied Foundation Models are One-Shot Learners — GEN-0 lays out the scaling-law + ossification-threshold half of Generalist’s bet, and GEN-1.5 (8+ months of continuous training on the same engine) is where the emergent-in-context-learning payoff shows up. Together the two blogs form the strongest single argument on the wiki for “pretraining engine scale is the lever, not recipe cleverness” — the same VLA Models recipe-lever position that GEN-1.5 stakes out but with the pretraining-side scaling curve made explicit. First filed instance of ossification in robotics — a phenomenon Hyperparameter scaling laws already tracks in the LLM literature (2102.01293 referenced by the post, How to Set the Learning Rate for Large-Scale Pre-training?) but not previously reported at the ~O(1B) scale the paper claims is needed for physical commonsense. Complements Dyna-2: A 1-Million-Hour Scaling Law for World-Action Models (Dyna-2’s million-hour scaling law on egocentric human video → robot transfer) with a real-robot manipulation substrate — same “clean power law from real interaction data” thesis on a different data axis. Complements Towards Machines with a Thousand Hands (GEN-1’s cross-end-effector scaling) as the depth-of-data companion to GEN-1’s breadth-of-end-effector story. The Harmonic Reasoning framing is an implicit critique of the VLA Models board’s System-1/System-2 recipes (Flexion’s Reflect, VoLo, RoboHarness) and of inference-time guidance patterns (πR², real-time chunking) — positions the async sensing/acting interleaving inside the model rather than as an orchestrator or a scheduling trick.