Robotics & Embodied Foundation Models: A Landscape Survey

Last verified: 2026-05-08. This document was originally drafted from training-data knowledge (January 2026 cutoff) and has since been verified and enriched against primary web sources (company blogs, arXiv, IEEE Spectrum, TechCrunch, Bloomberg, etc.). Inline [Source] citations link to verifying sources. Claims that could not be independently verified are flagged [unverified].

This document is a learning-oriented map of the embodied-foundation-model and humanoid-robotics landscape, organized for someone joining Applied Intuition's Data Intelligence team. The lens throughout is: where does the data come from, how is it labeled, and where does synthetic data fit?

---

A. The VLA (Vision-Language-Action) Thread

A VLA model takes pixels + a natural-language instruction and outputs robot actions (joint targets, end-effector deltas, or discrete tokens). The genre crystallized at Google in 2022–2023 and has since become the dominant recipe for "generalist" manipulation policies.

A.1 Google RT-1 → RT-2 → RT-X / Open X-Embodiment

RT-1, late 2022: 35M-parameter Transformer trained on ~130k teleop episodes collected over 17 months across 13 Everyday Robots arms. Tokenized 7-DoF arm + base actions. Established that scaling laws hold for manipulation given enough demos. RT-1.

RT-2, mid-2023: fused a pretrained VLM (PaLI-X / PaLM-E backbones) with action tokens. Key trick — cast actions as text tokens and co-fine-tune on web VQA + robot trajectories — let RT-2 inherit semantic generalization (e.g., "pick up the extinct animal" → dinosaur toy). RT-2.

Open X-Embodiment / RT-X, October 2023: a multi-institution dataset and model effort. The dataset aggregated ~1M+ real-robot trajectories from 22 embodiments across 21 institutions, pooling 60 pre-existing datasets covering 527 manipulation skills in a shared (RLDS) schema. RT-1-X and RT-2-X showed positive transfer across embodiments — a single policy improving on multiple lab-specific baselines. The field's biggest "ImageNet moment" claim for robotics. Open X-Embodiment paper (arXiv:2310.08864), project page.

A.2 OpenVLA (Stanford / Berkeley / TRI / MIT)

OpenVLA, June 2024 (arXiv:2406.09246): 7B-parameter open-source VLA on Llama-2-7B + DINOv2 + SigLIP visual encoders, trained on 970k OXE manipulation trajectories on a 64×A100 cluster for 15 days. Outperformed RT-2-X (55B) by +16.5% absolute task success rate despite being 7× smaller, and shipped weights + training code + LoRA fine-tuning recipes. The de-facto open baseline for VLA research. Authors span Stanford, UC Berkeley, TRI, Google DeepMind, Physical Intelligence, and MIT. OpenVLA project, HF weights.

A.3 Octo (Berkeley)

Octo, early 2024: a small (27M / 93M) Transformer trained from scratch on 800k OXE episodes with a diffusion-policy action head. Emphasizes flexible conditioning (language or goal images), modality dropout, and easy fine-tuning. Popular for academic groups that don't want a 7B VLM in the loop. Octo.

A.4 Physical Intelligence — π0, π0-FAST, π0.5

Physical Intelligence (PI) — the Karol Hausman / Sergey Levine / Chelsea Finn startup — has shipped three notable models.

• π0, paper posted October 31, 2024 (arXiv:2410.24164): flow-matching VLA on PaliGemma-3B + a separate "action expert" Transformer. Action chunks predicted via flow matching at 50 Hz. Trained on OXE + a large internal dexterous mobile manipulator dataset across multiple embodiments. Demos: laundry folding, table bussing, box assembly — long-horizon, contact-rich. Open-sourced via openpi GitHub. π0 blog.
• π0-FAST, weights released February 2025 (paper January 2025): replaces flow matching with frequency-space action tokenization ("FAST" tokens, based on the Discrete Cosine Transform), making the policy purely autoregressive — easier to train with standard VLM infra and up to 5× faster to train. π0-FAST research page.
• π0.5, April 22, 2025 (arXiv:2504.16054): the "open-world" version, co-trained on heterogeneous data — web data + cross-embodiment robot data + high-level semantic prediction + verbal supervision, aimed at generalization to unseen homes. Demoed cleaning kitchens and bedrooms in houses the robot had never seen during training — arguably the strongest public evidence to date that VLA generalization extends out-of-distribution. π0.5 blog.
• π0.6, model card November 17, 2025 (PI06 model card PDF) — the latest in the family at time of verification.

What's distinctive about PI: flow-matching action heads at high frequency, deliberate cross-embodiment training (including bimanual mobile bases), open weights, and emphasis on long-horizon dexterous tasks rather than tabletop pick-and-place.

PI funding (verified): $70M seed (March 2024) → $400M Series A at $2.4B valuation (November 2024, Bezos / OpenAI / Thrive / Lux / Bond) → $600M Series B at $5.6B post-money (November 2025, led by CapitalG — Bloomberg, The Robot Report). A reported $1B round at ~$11B valuation is in talks as of March 2026 [unverified — reported, not closed].

A.5 Figure — Helix / Helix 02

Helix, February 20, 2025: Figure's in-house humanoid VLA, a two-system stack. "System 2" — an internet-pretrained VLM running at 7–9 Hz reasoning over scene + instruction. "System 1" — an 80M visuomotor Transformer at 200 Hz outputting continuous control of the entire upper body (wrists, torso, head, individual fingers). They communicate via a latent vector. Demoed two humanoids collaboratively putting groceries away from a single neural net. Runs entirely onboard low-power embedded GPUs. Trained on internal teleop data; full specifics unpublished. Figure Helix.

Helix 02 (announced January 27, 2026): adds full-body autonomy (locomotion + manipulation + balance as one continuous learned system) and introduces a "System 0" foundation layer that runs balance/coordination at kilohertz rates underneath System 1/2. Demoed end-to-end autonomous dishwasher loading and living-room cleanup. Figure Helix 02.

A.6 1X Technologies — World Model + Neural Stack

1X's NEO humanoid runs an end-to-end neural net. Their distinguishing public artifact is the 1X World Model (late 2024): an action-conditioned generative video model used for both evaluation (compare predicted vs. real video) and data augmentation. Trained on thousands of hours of EVE/NEO interaction video from homes via teleop and the employee fleet. NEO Beta was unveiled August 30, 2024, and the consumer NEO launched in October 2025 at a $20,000 price point — though early reporting noted that initial deployments rely heavily on remote human teleoperation rather than full autonomy (dronexl coverage). 1X raised a $100M Series B in January 2024 (EQT Ventures). 1X World Model.

A.7 DeepMind — RT-2-X, Gemini Robotics

RT-2-X was DeepMind's contribution to the Open X-Embodiment effort (RT-2 fine-tuned on the cross-embodiment mix).

Gemini Robotics & Gemini Robotics-ER, announced March 2025 (DeepMind blog): Gemini-2.0-based VLA models. The family has two members: a "Robotics" action model that emits robot actions, and an "Embodied Reasoning (ER)" model that emits 3D bounding boxes, grasps, trajectories, and code — a foundation for downstream policies. DeepMind partnered with Apptronik for humanoid integration and with other OEMs (Boston Dynamics, Agility, Agile Robots, Enchanted Tools) as trusted testers.

Subsequent releases:

• Gemini Robotics On-Device (June 24, 2025): a VLA optimized to run locally on the robot, adapts to new tasks with as few as 50–100 demos. Ships with a Gemini Robotics SDK. On-Device blog.
• Gemini Robotics 1.5 + Gemini Robotics-ER 1.5 (September 2025): the "thinking" generation. ER 1.5 is a high-level reasoner that orchestrates multi-step plans and hands off to the 1.5 action model. Cross-embodiment transfer demonstrated across ALOHA-2, Franka bi-arm, and Apptronik Apollo without retraining. Gemini Robotics 1.5 blog.
• Gemini Robotics-ER 1.6 (DeepMind blog) — incremental upgrade to the ER reasoner.

A.8 Tesla Optimus

Public information on Optimus's stack is fragmentary. Known/confirmed: Optimus uses an end-to-end neural net for hand control and walking, shares perception infrastructure with the FSD vision stack (HydraNets, occupancy networks), and is trained partially on mocap suits + VR teleop at Tesla's Fremont facility. Tesla has not published a formal VLA paper. Gen 2 launched late 2023; Gen 3 was teased on Tesla's Q3 2025 earnings call with a target unveil in Q1 2026 [unverified production-ready date]. Reported Gen 3 specs: ~173 cm, ~57 kg, FSD-v13 end-to-end NN, 22-DoF tendon-driven hands (up from 11 in Gen 2), 2.3 kWh battery, target retail $20–25k. Performance in many public demos has been disclosed as teleoperated (e.g., the October 2024 "We, Robot" event). Treat capability claims with skepticism until independently demonstrated. Tesla AI.

A.9 Skild AI — Skild Brain

Skild AI (Deepak Pathak, Abhinav Gupta — CMU spinout) raised a $300M Series A on July 9, 2024 at a $1.5B valuation (Lightspeed, Coatue, SoftBank, Bezos Expeditions — Skild blog), followed by a ~$500M SoftBank-led raise in April 2025, and a ~$1.4B SoftBank-led round in January 2026 at a $14B valuation (Crunchbase News, TechCrunch). Their pitch is a single "general-purpose robot brain" ("Skild Brain") that generalizes across hardware: quadrupeds, manipulators, humanoids. They claim to have trained on a dataset "1000× larger than competitors" combining real, sim, and human video — specifics not disclosed. Few public demos beyond promo videos. Capability claims remain largely unverified. Skild AI.

A.10 Covariant — RFM-1 (now Amazon)

Covariant built warehouse pick-and-place arms running on a multimodal foundation model called RFM-1 (Robotics Foundation Model 1), announced March 11, 2024. RFM-1 ingests text, images, video, robot states, and actions as a unified token stream — an autoregressive 8B model trained on years of in-warehouse pick data plus internet video. On August 30, 2024, Amazon hired the three Covariant founders (Pieter Abbeel, Peter Chen, Rocky Duan) plus ~25% of staff, and signed a non-exclusive license to Covariant's models — a "reverse acquihire" structure (TechCrunch, GeekWire). A 2025 whistleblower disclosure put the deal value at ~$380M plus a $20M follow-on licensing fee — below Covariant's prior $625M valuation. Ted Stinson now leads the residual Covariant. Covariant RFM-1.

A.11 Apptronik + Apollo (DeepMind partnership)

Apptronik's humanoid Apollo (a 1.6m, 73kg general-purpose humanoid) entered commercial pilots starting in 2024. The Mercedes-Benz agreement (announced March 2024 — PR Newswire) is testing Apollo at Mercedes' Berlin-Marienfelde Digital Factory and a Hungarian site for assembly-kit delivery; Mercedes also took a low-double-digit million-euro stake. GXO launched a multi-phase Apollo R&D in 2024 (lab → distribution-center deployment). Both remain pilot/R&D phase, not full commercial production deployments. In December 2024, Apptronik announced a collaboration with Google DeepMind to build foundation models for Apollo, leveraging Gemini Robotics. Funding: $350M Series A in February 2025 (co-led by B Capital and Capital Factory, with Google as a participant — not lead — per TechCrunch); a Series A extension in February 2026 brought the total to $935M at a ~$5B valuation (CNBC). Apptronik + DeepMind.

---

B. Humanoid Hardware Companies & Data Strategies

Pattern. The humanoid market split into three camps: (1) vertically integrated AI + hardware (Figure, 1X, Tesla, XPENG), (2) hardware + foundation-model partnership (Apptronik–DeepMind, BD–TRI), and (3) hardware-only / research platforms (Unitree, Fourier). The first formal commercial humanoid RaaS contract — Agility's GXO Logistics deal at the SPANX Atlanta DC (June 27, 2024) — remains the strongest empirical anchor for "humanoids actually working" in real environments, though scope is narrow (tote handoffs to conveyors). Most "deployment" announcements in 2024–25 are still pilots, not steady-state commercial production: BMW publicly retired the Figure 02 pilot in November 2025 with no current robots on its line, and Mercedes' Apptronik partnership has not graduated past pilot.

---

C. Manipulation, Mobile, and Locomotion Datasets

C.1 Open X-Embodiment (OXE)

The biggest aggregated robot dataset to date. The 2023 release pooled >1M trajectories from 60 datasets across 22 embodiments and 21 institutions, covering 527 manipulation skills [verified: arXiv:2310.08864]. Subsequent expansions have appeared (e.g., OXE-AugE in 2025 augments the OXE seed to >4.4M trajectories via robot augmentation), but the original ~1M figure is the headline; the "~2.4M" oft-cited "OXE 1.5" figure is unverified [unverified]. Standardized to a common RLDS schema. Limitations: very long tail (a few datasets dominate, many embodiments have a handful of episodes), and mostly tabletop manipulation. OXE project, GitHub.

C.2 DROID

DROID, March 2024 (arXiv:2403.12945): 76k successful demonstration trajectories / ~350 hours / 564 scenes / 52 buildings / 86 tasks, collected by 50 data collectors across 13 institutions on 3 continents over 12 months. Hardware: Franka Panda 7-DoF + 2 adjustable Zed 2 stereo cameras + wrist-mounted Zed Mini + Oculus Quest 2 teleop. Unlike OXE's heterogeneity, DROID is one platform / one teleop interface — much cleaner for training and evaluation. Released CC-BY 4.0 with code and pretrained checkpoints. DROID project.

C.3 Bridge / BridgeData V2

BridgeData V2 (Berkeley): ~60k demos across 24 environments on a low-cost WidowX arm, 2022–2023. Designed for cheap reproducibility, used heavily in academic VLA papers. Bridge V2.

C.4 RoboNet

2019, ~15M frames across 7 robot platforms — historically important as the first cross-embodiment effort, largely superseded by OXE.

C.5 MimicGen / Synthetic Demos

MimicGen (Nvidia / UTexas, 2023) generates new demos from a small seed by adapting reference trajectories to new object poses in simulation — ~10 source demos → 1000s of synthetic demos. DexMimicGen (2024) extends to bimanual dexterous tasks. The practical answer to "we can't afford 100k human demos per task." MimicGen.

C.6 BEHAVIOR-1K, RoboCasa, RoboHive, Habitat

• BEHAVIOR-1K (Stanford, Fei-Fei Li) — 1000 long-horizon household activities in Omnigibson, grounded in human surveys of "what should a home robot do."
• RoboCasa (Nvidia / UTexas, 2024) — Robosuite-based sim framework with 100+ kitchen layouts and procedurally generated objects, built for VLA scaling.
• RoboHive — meta-benchmark unifying many manipulation envs under one API.
• Habitat (Meta) — navigation + embodied AI; Habitat 3.0 (2023) added humanoid avatars for HRI research.

C.7 Industrial / Warehouse Data

Most warehouse-scale data is proprietary: Amazon Robotics, Symbotic, Ocado, Covariant (now Amazon), GXO all run live picks at billions-per-year scale but do not publish. Public warehouse benchmarks remain weak — a real opportunity area.

---

D. Sim-to-Real & Data Augmentation for Robots

D.1 Domain Randomization (Classic)

OpenAI's 2017–2019 work on robotic hands (Rubik's cube) established the playbook: train in sim with randomized physics, lighting, textures, and camera poses; the real world becomes "just another sample." Still the workhorse for locomotion (e.g., ANYmal, Spot, Unitree controllers all trained this way using Isaac Gym / Mujoco MJX in massive parallel envs).

D.2 Generative Video as World Model / Data Engine

A 2024–2025 wave: rather than randomize physics, generate plausible futures with video models.

• 1X World Model (2024) — action-conditioned video generation, used for both eval and training data. 1X.
• Genie / Genie 2 / Genie 3 (DeepMind) — interactive world models. Genie 2 (Dec 2024) generated 3D-consistent rollouts from a single image. Genie 3 (announced Aug 5, 2025 — DeepMind blog) is the first real-time interactive general-purpose world model: text-prompted dynamic worlds at 24 fps, 720p, with multi-minute consistency through emergent memory. Initially research-preview; February 2026 rolled into Project Genie for Google AI Ultra users in the US.
• RoboGen (CMU / MIT, 2023) — uses LLMs to propose tasks, scenes, and reward functions, then fills in 3D assets and trains policies in sim.
• UniSim / Sora-style approaches — universal action-conditioned video models as a substitute for hand-built simulators. Still research-stage; the open question is whether they're physics-accurate enough for contact-rich manipulation.

D.3 Nvidia Stack — Isaac Sim, Isaac Lab, GR00T

This is now the dominant simulation stack outside of academic Mujoco.

• Isaac Sim — GPU-accelerated physics on PhysX, photoreal rendering on Omniverse. Workhorse for industrial digital twins.
• Isaac Lab — open-source RL training framework on top of Isaac Sim (replaces older Isaac Gym). Supports massively parallel envs (~thousands in one GPU) for locomotion and manipulation.
• Project GR00T — Nvidia's "general-purpose humanoid foundation model" initiative announced at GTC March 2024.
  • GR00T N1 (announced March 18, 2025 at GTC, Nvidia press, arXiv:2503.14734) — first open release. Dual-system architecture (slow VLM "System 2" + fast diffusion-transformer action head "System 1"), trained on a mix of real teleop, OXE, human video (egocentric), and Isaac-generated synthetic. Open weights.
  • GR00T N1.5 (June 2025, demoed live at automatica 2025 on Franka FR3 dual-arm — Franka news) — substantial upgrade: 38.3% success rate on 12 DreamGen tasks vs 13.1% for N1. GR00T N1.5 research page.
  • GR00T N2 — not announced as of verification date (May 2026). Earlier "rumored" framing was incorrect; no GR00T N2 has been officially released. [unverified — drop or downgrade].
• GR00T-Mimic (announced GTC 2025) — Nvidia's pipeline for generating synthetic robot demonstrations from a handful of human demos using a combo of MimicGen-style replay + Cosmos generative video augmentation.
• GR00T-Gen (announced GTC 2025) — generative pipeline for scenes, layouts, and tasks in Isaac (procedural homes, warehouses) — analogous to RoboCasa but tied into the Nvidia stack.
• Cosmos (announced Jan 6, 2025 at CES, Nvidia press) — Nvidia's family of "world foundation models" (diffusion + autoregressive variants) released under an open model license. Trained on 9,000 trillion tokens drawn from 20M hours of real-world human, environment, industrial, robotics, and driving data. Three tiers: Nano (edge/low-latency), Super (baseline), Ultra (max quality, distillation source). Early adopters: 1X, Agile Robots, Agility, Figure, Foretellix, Fourier, Galbot, Hillbot, IntBot, Neura Robotics, Skild AI, Virtual Incision, Waabi, XPENG, Uber. Nvidia Cosmos.

D.4 3DGS / NeRF Capture for Sim

Increasingly common: scan a real scene with a phone or capture rig, reconstruct as a 3D Gaussian Splatting or NeRF asset, drop into Isaac/Mujoco for rendering. Robot Studios like PolyCam, Luma, and research projects like GauSim and SplatSim have made this a near-commodity workflow. The hard part is recovering accurate geometry + material + collision — rendering is great, physics is still mostly a hand crank.

D.5 Real-to-Sim Digital Twins

• RoboCasa (UTexas / Nvidia) — 100+ kitchen scenes, procedurally generated, built specifically for VLA scaling experiments.
• PolyGen / Genesis — Genesis (CMU 2024) is a unified differentiable physics + rendering engine that pitches itself as a generative simulator.
• Industry: Applied Intuition's own simulator stack, Nvidia Drive Sim, dSPACE, etc., for AVs; analogous "robot digital twin" tooling is much less mature. This is one of the cleanest whitespace areas.

---

E. Whitespace and Bottlenecks

E.1 The Data Problem: No Internet for Robots

LLMs scaled because the internet is the dataset. Robotics has no equivalent. The OXE dataset is ~2M trajectories — orders of magnitude smaller than what's needed to match LLM-style emergence. The community's three workarounds are:

1. Teleoperation at scale. Cost is the binding constraint. A skilled teleoperator collects roughly 50–150 demos per hour per robot. At ~$30/hr fully loaded, 1M demos costs $250k–$1M and many robot-years. Companies like Figure, PI, and 1X are all running private teleop farms in this range.
2. Egocentric human video. Project Aria (Meta), Ego4D (3,670 hours), Ego-Exo4D (Meta + 15 universities, 2024) and HoloAssist all aim to harvest "what humans do with their hands" as a free supervisory signal. The catch: no action labels, no force feedback, no teleop calibration. Bridging the human-to-robot gap requires retargeting + sim or learned correspondence.
3. Synthetic / generative. MimicGen-style replay, GR00T-Mimic, and video world models. The honest answer is: we don't yet know how much of the gap synthetic can close for contact-rich manipulation.

E.2 Teleop Economics

The economics matter for whitespace-spotting. If you assume:

• ~$50/hr loaded cost per teleoperator (incl. equipment)
• 100 demos/hr good demos
• Need ~10k demos per task to be reliable, ~100 tasks → 1M demos
• → ~$500k of pure teleop cost per "general-purpose" model, before failure/cleanup overhead

Scaling 10x gets prohibitive fast. This is why synthetic-data multipliers (50x–1000x from each real demo) and policy-conditioned scenario generation are existential.

E.3 Generalization Across Embodiments

OXE / RT-X showed positive transfer across similar embodiments (mostly 6–7 DoF arms with parallel grippers). Transfer to dexterous hands, mobile bases, humanoids is much weaker — the action spaces are too different. Cross-embodiment is partially solved for arms, mostly unsolved for legs and hands. The bet of Skild and PI is that with enough scale + the right tokenization, this works. Verdict pending.

E.4 Multi-Task vs Specialist

Empirically, in 2024–2025, specialist policies beat generalists on the specialist's task, but generalists win on novel tasks and on cost-of-ownership. This is the same pattern that played out in NLP from 2018–2022. The economics will likely favor a small number of generalist VLAs per industry vertical, fine-tuned per customer.

E.5 Evaluation Is Broken

This is the area where Applied Intuition–style tooling is most needed.

• No standard benchmark for generalist policies. Each lab reports on its own real-robot tasks, often with non-reproducible setups. RoboArena / RoboHive / SimplerEnv try to standardize sim eval, but sim-real gap means sim numbers don't predict real performance well.
• Real robot eval is expensive. A serious eval is dozens of trials across many tasks with human raters. Companies do this internally but rarely publish.
• Long-horizon eval is even worse. No agreed metric for "how good was this 5-minute laundry-folding rollout?" Companies use partial-credit rubrics, time-to-completion, or task-success-at-step-N — all noisy.
• Coverage / scenario generation. Like AVs in 2018–2020, the field needs to move from "pass rate on a fixed test set" to "what's our coverage of meaningful scenarios?" This is exactly the AV-validation pattern Applied Intuition built its business on.

---

F. Implications for Applied Intuition / Data Intelligence

The thesis: the AV-data tooling pattern transfers to robotics, but is much less mature and more fragmented. Concrete opportunity areas:

F.1 Labeling Robot Demos

Teleop episodes need cleaning (failed demos, off-policy moments, gripper miscalibrations). Multi-modal alignment (video + force-torque + proprioception) is non-trivial. VLAs need hierarchical language labels (goal → sub-skill → primitive); most labs do this ad-hoc. Embodiment normalization — mapping multiple robots into a canonical action schema — remains rough (OXE's schema is a starting point but limited).

F.2 Scenario Coverage for Manipulation

AV analog: "tested left-turn-on-red in fog?" Robot analog: "tested grasp-from-cluttered-shelf with reflective objects under low light?" Requires scene parametrization (objects, layouts, lighting, distractors) plus a coverage metric over that space, tied to scenario generation — procedural (RoboCasa, GR00T-Gen) or generative (Cosmos / video world models).

F.3 Policy Validation in Sim

Run a candidate policy against a battery of scenarios and surface regressions. Catch sim-real gap with dual eval (same policy in sim + on a small real-robot fleet; learn the residual). Replay-based validation as a sanity check.

F.4 Data Curation

Active learning over teleop pools (which demos are most informative?), difficulty / diversity scoring (avoid yet-another-pick-from-table), de-duplication, drift detection, embodiment-specific filtering.

F.5 Synthetic Generation as a Product

MimicGen-style replay augmentation (10x–1000x per real demo), generative video augmentation (Cosmos / 1X-WM style) for visual robustness, procedural scene generation for sim-based RL. Most labs roll these tools themselves, badly. Productizing them with metadata, lineage, and policy-impact tracking is a clear gap.

F.6 Long-Horizon Eval Tooling

Rubric-based human eval pipelines (RLHF tooling, but for episodes), task graph decomposition with per-step success scoring, failure mining (cluster by mode — perception miss, slip, planning, recovery — and feed back into the curriculum).

Bottom line for Data Intelligence. The primitives that matured for AVs in 2017–2024 — scenario libraries, coverage metrics, synthetic generation tied to scenario gaps, sim-real transfer harnesses, data curation pipelines — are now in early demand for robotics. The field is roughly 5–7 years behind AVs in tooling maturity but moving faster because the foundation-model recipe transfers cleanly from LLMs. The dirty work of mapping kinematics, sensor configs, and action spaces across robots is currently per-paper; a canonical embodiment-translation layer + dataset registry is overdue.

---

Sources

Models & papers

• RT-1 — Robotics Transformer 1 (Google, 2022)
• RT-2 (Google DeepMind, 2023)
• Open X-Embodiment / RT-X paper (arXiv:2310.08864) · project page · GitHub
• OpenVLA (arXiv:2406.09246) · project page · HF weights
• Octo (Berkeley, 2024)
• π0 paper (arXiv:2410.24164) · PI blog
• π0-FAST research page
• π0.5 paper (arXiv:2504.16054) · PI blog
• openpi GitHub (PI open weights)
• Figure Helix · Figure Helix 02
• 1X World Model
• Gemini Robotics announcement (DeepMind, March 2025)
• Gemini Robotics On-Device (June 2025)
• Gemini Robotics 1.5 (Sept 2025)
• Gemini Robotics-ER 1.6
• Apptronik + Google DeepMind
• Covariant RFM-1 · Amazon-Covariant deal (TechCrunch)
• Skild AI homepage · Skild $14B raise (TechCrunch, Jan 2026)
• Tesla AI
• Boston Dynamics Atlas · BD–TRI partnership (TechCrunch, Oct 2024)
• Agility Robotics – GXO RaaS announcement (June 27, 2024)
• Sanctuary AI – Magna pilot (April 2024)

Datasets

• DROID paper (arXiv:2403.12945) · project
• BridgeData V2
• MimicGen
• BEHAVIOR-1K / Stanford
• RoboCasa
• Ego4D / Ego-Exo4D (Meta)

Sim & generative stack

• Nvidia Project GR00T / GR00T N1 announcement (March 2025) · GR00T N1 paper (arXiv:2503.14734) · GR00T N1.5 page
• Nvidia Cosmos announcement (Jan 2025) · product page
• Nvidia Isaac Lab / Sim
• DeepMind Genie 3 (Aug 2025)
• DeepMind Genie 2 (Dec 2024)

Funding / business news

• Figure Series C $1B+ at $39B (Sept 2025) · Figure BotQ
• BMW – Figure Spartanburg pilot results (BMW press, Nov 2025) · WardsAuto coverage
• Apptronik $350M Series A (TechCrunch, Feb 2025) · Apptronik $520M extension (CNBC, Feb 2026)
• Physical Intelligence $600M / $5.6B (Bloomberg, Nov 2025)
• Mercedes-Benz – Apptronik commercial agreement (March 2024)
• XPENG AI Day 2025 — next-gen Iron + VLA 2.0
• Unitree G1 launch (The Robot Report)