Frontier labs & startups
17Companies whose entire business is getting manipulation to work outside the lab. This is where the sharpest data/compute/results disclosures live, because reliability is their product.
Sanctuary AI Unveils Phoenix
"Carbon" cognitive architecture — a symbolic + LLM + RL hybrid, not a pure end-to-end policy. 20-DOF hands with tactile sensing.
Covariant Introduces RFM-1
8B-param any-to-any sequence model over text/image/video/robot-action/measurement tokens, trained on years of real warehouse deployment data from Covariant Brain — one of the few reports built on genuine multi-year production data rather than a curated collection sprint.
π0: A Vision-Language-Action Flow Model for General Robot Control
Tech report. 7 robot embodiments, 68 tasks, PaliGemma-3B VLM backbone + flow-matching action head. The reference architecture most later VLA work positions itself against.
Helix: A Vision-Language-Action Model for Generalist Humanoid Control
Dual-system architecture: 7B VLM at 7–9Hz feeding an 80M-param visuomotor policy at 200Hz. First VLA controlling a full humanoid upper body including individual fingers; demos two robots collaborating on one task.
Helix Accelerating Real-World Logistics
Deployment update moving Helix from demo into logistics task settings.
Welcome to Robot Park
90,000 sq ft dedicated data-collection/training facility for Apollo; partnerships with Google DeepMind (Gemini Robotics) and NVIDIA (GR00T) named directly.
DYNA-1 Research
First commercial dexterous foundation model claim. 24-hour autonomous run, 900+ napkins folded at 99% success in a live deployment (hotels, restaurants, laundromats), not a lab benchmark.
π0.5: a VLA Model with Open-World Generalization
Blog + paper (arXiv 2504.16054). "Knowledge insulation" training; cleans entirely unseen homes; 10–15 min multi-stage tasks.
openpi (GitHub)
Open-source code and weights for π0 and π0-FAST — the rare case of a frontier lab shipping runnable weights, not just a paper.
Building the General-Purpose Robotic Brain
Hierarchical high-frequency/low-frequency policy architecture; the "omni-bodied brain" thesis — one brain, many robot bodies.
One Model, Any Scenario: End-to-End Locomotion from Vision
Locomotion-specific generalization results — one policy across terrains/scenarios without per-scenario retraining.
The Case for an Omni-Bodied Robot Brain
Architecture philosophy — trained across 100,000 different simulated robot bodies to force embodiment-general representations.
Learning by Watching Human Videos
Human-video-to-affordance extraction approach — sidesteps teleop bottleneck by mining action signal from unstructured human footage.
From Video to Action: A New Way Robots Learn
Primary source for the 1X World Model (1XWM). 14B-param video model + inverse dynamics model, trained on 900 hrs egocentric human video + 70 hrs NEO robot data + 400 hrs unfiltered play. Positioned explicitly against direct-action VLA approaches.
1X World Model: Evaluating Bits, Not Atoms
The fuller technical PDF behind the 1XWM blog post above — evaluation methodology for judging a world model by prediction quality rather than physical rollout alone.
Introducing S1: In-Context Learning for Robotics
One-shot, video-prompted task learning — no fine-tuning. Claims a single demo is worth roughly 380 post-training episodes, and a 7× gain on unseen tasks over language-prompted baselines at 100k-hour pretraining scale.
Dyna Robotics Unveils Dyna-2 World Action Model
Press release. Pretrained on 1M+ hours of egocentric human video; reports manufacturing success jumping 20%→80–90% from pretraining scale alone; 87% zero-shot pass rate vs. 46% for DYNA-1 (1.55×); 133% gain on instruction-following.
Big-tech robotics efforts
8Larger research orgs with compute and data scale most startups can't match — useful for seeing what changes when the constraint shifts from capital to something else.
RT-2: Vision-Language-Action Models
Blog. 44%→62% manipulation success jump over RT-1 by co-training on web-scale vision-language data alongside robot trajectories.
Open X-Embodiment / RT-X
Project site + paper. 22 robots, 21 institutions, 1M+ demonstrations, 527 skills. Showed cross-embodiment co-training beats training each robot alone — the dataset most later generalist policies pretrain on.
Genie 2: A Large-Scale Foundation World Model
Generates playable, action-controllable 3D environments from a single image — pitched as training/eval substrate for embodied agents rather than a manipulation policy itself. No full paper published, unlike Genie 1.
GR00T N1
Dual-system model: VLM at 10Hz + diffusion-transformer action head. Trained on human video + real/sim robot trajectories + synthetic data; pretraining used up to 1,024 GPUs (~50,000 H100-hours for the 2B model).
Gemini Robotics Brings AI into the Physical World
Gemini Robotics + Gemini Robotics-ER, trained primarily on ALOHA 2 bimanual data.
Gemini Robotics 2: Whole-Body Intelligence
Whole-body control update to the Gemini Robotics line.
Building a Synthetic Motion Generation Pipeline for Humanoid Robot Learning
780K synthetic trajectories generated in 11 hours (≈6,500 human-demo-equivalent hours); reports +40% GR00T N1 performance when synthetic data is combined with real data. The clearest sim-scale-vs-real-scale tradeoff number in this whole collection.
How to Evaluate General-Purpose Robot Policies for Real-World Deployment
Robot-agnostic benchmarking methodology: graded partial-credit task scores, a trajectory-smoothness metric (SPARC), structured failure-event logging — an attempt to measure reliability, not just success rate.
Open-source generalist policies
14Public architectures you can actually read the code for and, in several cases, run yourself. There are hundreds of VLA variants in the literature at this point — this is a curated cut of the ones with real released code/weights and a notable architectural idea or result, not an exhaustive list.
Diffusion Policy
Action-as-denoising-diffusion — the paper that solved the multimodal-action / mode-averaging problem underlying most modern action heads, including π0's and GR00T's. RSS 2023, foundational.
RoboCat
Self-improving generalist agent: visual-goal-conditioned decision transformer. Adapts to new embodiments from 100–1000 examples, then bootstraps more of its own training data — a real (if early) self-improvement loop.
RDT-1B: Robotics Diffusion Transformer
1.2B-param diffusion transformer — largest bimanual manipulation foundation model at release. Pretrained on 46 datasets / 1M+ episodes plus 6k ALOHA episodes; fine-tunes to new skills from 1–5 demos. Code.
CogACT
Componentized VLA with a separate dedicated action module. Beats OpenVLA at the same 7B size by 35% in sim / 55% on real robots, and beats the 55B RT-2-X by 18% in sim — a strong data point that action-head design matters as much as backbone scale. Code + weights.
RoboVLM(s)
A "what actually matters" framework paper for building VLAs — fully open code, models, datasets, and toolkits, not just a single checkpoint. SOTA in sim and real-world with deliberately minimal manual design choices. Project page.
SpatialVLA
Trained on 1.1M real robot episodes; introduces Ego3D Position Encoding and Adaptive Action Grids for spatial transfer across different robot embodiments. Zero-shot deployable. RSS 2025. Code + weights.
OpenVLA-OFT
An optimized fine-tuning recipe for OpenVLA — parallel decoding, action chunking, continuous actions. Pushes success from 76.5% to 97.1% while running 26× faster; works on bimanual ALOHA. Code + weights.
RLDX-1 Technical Report
75+ authors incl. Dongyoung Kim, Huiwon Jang (RLWRLD). Introduces the Multi-Stream Action Transformer (MSAT) — unifies vision, tactile sensing, high-DoF actuation, and memory-aware decision-making in one architecture, explicitly targeting VLA gaps around motion awareness, long-horizon memory, and physical sensing. Product page: rlwrld.ai/en/rldx-1 · Hugging Face weights.
Alignment Unlocks Scale for Robotic Manipulation Foundation Models
Trained on ~38,100 hours of pretraining data drawn from open datasets plus human video; reports zero-shot instruction following, error recovery, and cross-embodiment transfer.
Datasets & data collection infrastructure
14Everyone eventually says the same thing: data, not architecture, is the bottleneck. This is the infrastructure built to attack that — new hardware for collecting demos faster, synthetic generation to multiply demos already collected, and the datasets that resulted.
RoboTurk
Crowdsourced remote teleoperation — 111+ hours across 54 users on 3 tasks in one week. Historically the first serious attempt at crowdsourced scale, worth reading to see how far the field has come in 6 years.
ACT / ALOHA
Low-cost (~$20k) bimanual hardware + action-chunking transformer. 80–90% success on fine-grained tasks from only ~10 minutes of demonstration each — the paper that made cheap high-quality teleop data collection mainstream.
BridgeData V2
60,096 trajectories, 24 environments, 13 skills on low-cost hardware, goal-image or language conditioned.
Open X-Embodiment (paper)
Full dataset and methodology behind RT-X, cross-listed here since it's as much a data-infrastructure story as a model story.
Mobile ALOHA
Adds a mobile base and whole-body teleop ($32k system). Co-training with static-ALOHA data pushes success to ~90% from only 50 demos per task.
UMI: Universal Manipulation Interface
Hand-held gripper for collecting demos from humans directly — no robot needed during data collection at all. Produces hardware-agnostic, deployable policies; zero-shot on dynamic, bimanual, long-horizon skills.
DROID
76k trajectories / 350 hours, 564 scenes, 84 tasks, collected by 50 people across 3 continents over 12 months. The "in-the-wild diversity beats controlled-lab diversity" dataset.
DexHub and DART: Towards Internet-Scale Robot Data Collection
Cloud-hosted AR teleoperation; reports 2.1× faster data-collection throughput than existing real-world teleop platforms, with lower physical fatigue on the operator.
DexMimicGen
Extends MimicGen to bimanual dexterous/humanoid hands — synthesizes large training sets automatically from a handful of human demos via scene/object randomization. MimicGen code. ICRA 2025.
AgiBot World Colosseo
1,001,552 trajectories, 2,976.4 hours, 217 tasks, 87 skills, 106 scenes, collected via 100 dual-arm humanoid robots with human-in-the-loop verification. Introduces the GO-1 policy and shows predictable scaling with data volume — reports +30% over OXE pretraining alone. IROS 2025 Best Paper Award finalist. Code.
Expanding Our Data Engine for Physical AI
100,000+ production hours collected across an SF lab plus global contributors. States the bottleneck bluntly: DROID + OXE combined amount to only ~5,000 hours — a useful anchor for how small even the "big" academic datasets are next to what a data-vendor considers a real training run.
Getting a Grip on Robotic Data Collection
Boston Dynamics-affiliated research institute's handheld/wearable data-collection tooling, in the same spirit as UMI.
Unfolding Robotics: Open-Source Shirt Folding from Data to Deployment
The thorough open-source empirical study you asked about by name. Full recipe — data collection with 10 human teleoperators on OpenArms, model, dataset, and deployment code — documented step by step. LeRobot repo.
LeRobot Hackathon Winners
Independent replication of shirt-folding: a hackathon team reached ~70% success with only 86 demos, comparing π0, ACT, and SmolVLA head to head — useful as a small-data, small-team data point against the big-lab numbers above.
RL fine-tuning & sim-to-real
7Where imitation learning is explicitly not the whole answer — real-world reinforcement learning after or instead of behavior cloning, and using simulation to multiply real-world experience.
HIL-SERL
Human-in-the-loop sample-efficient RL. Reports 100% success on dexterous real-world tasks after just 1–2.5 hours of training, beating imitation baselines outright. Paper (also in Science Robotics) · code · LeRobot integration.
SERL
The software suite HIL-SERL builds on — sample-efficient real-world robotic RL as a reusable stack, not a one-off result.
RLDG
Distills generalist VLA policies using RL rather than pure imitation distillation.
Develop Humanoid Robot Policies End-to-End with Isaac GR00T
The full Isaac Lab / Isaac Sim pipeline connecting synthetic generation to GR00T training — pairs with the synthetic-motion pipeline post above.
RL-100
Real-world RL manipulation results at a higher performance tier than earlier RL-from-scratch attempts.
WMPO: World Model-Based Policy Optimization for VLA Models
Uses a learned world model to optimize a VLA policy — a bridge between the RL and world-model buckets.
Learning While Deploying: Fleet-Scale RL for Generalist Robot Policies
RL applied across an entire deployed fleet rather than one robot at a time — closer to how a production company would actually have to run this.
World models
14The bet that predicting future video/state, then deriving actions, generalizes further than predicting actions directly. A handful of these are demonstrated on real robot manipulation; others are general-purpose video/world models included for lineage even though their manipulation link is weaker — noted where that's the case.
DreamerV3
General world-model-based RL algorithm — learns a world model, then trains a policy purely on imagined rollouts inside it. Demonstrated on real-world robot arm cube manipulation without task-specific tuning, alongside dozens of other domains. Project page.
UniSim: Learning Interactive Real-World Simulators
Trained on real + simulated robot data plus internet video to build an action-conditioned video simulator, then used to train real-world-deployable robot policies entirely inside the simulator — a direct closing of the sim-to-real loop via a world model. Project page.
Genie: Generative Interactive Environments
The foundational Genie paper — trained on internet video, generates playable 2D environments from a single image. Establishes the lineage behind Genie 2 and 3 below; weak direct link to robot manipulation on its own.
V-JEPA
Predicts in learned representation space rather than pixel space — LeCun's world-model thesis in its first robotics-adjacent form. Foundational for V-JEPA 2 below; not itself deployed on a real robot.
RoboDreamer
Compositional video-diffusion world model for manipulation planning — generates imagined videos of an action sequence to guide a real robot policy before it acts. Project page.
iVideoGPT
Scalable autoregressive world-model framework; action-conditioned video prediction feeds model-based RL evaluated directly on real robot arm manipulation tasks.
Genie 2
See full entry under Big Tech above — listed here too since it's the other pole of the world-model bet.
Cosmos World Foundation Models
World Foundation Models trained on 20M+ hours of driving and robotics video, purpose-built for physical AI. Used directly to generate synthetic training data for GR00T and other manipulation policies — one of the strongest manipulation links in this category. Paper.
WorldVLA
Explicitly unifies a world model and a VLA action model in one autoregressive framework — the world model predicts future visual states to sharpen the VLA's action generation. Evaluated on LIBERO; directly bridges this category with Open-Source Policies and the WMPO entry above.
Genie 3
Real-time interactive world model generating consistent, 3D-navigable environments at 24fps/720p with minutes-long consistency. Pitched as "a foundation for the physical world" but general-purpose (games/simulation) rather than manipulation-specific — the biggest scale jump in the Genie line, weak direct robotics link so far.
Marble
Generates persistent, explorable 3D world models from images or text. Positioned for spatial intelligence broadly; robotics-manipulation application is implied/future-facing rather than demonstrated yet — included for the direction, not a result.
Redwood AI World Model
Related post on 1X's world-model line, adjacent to the flagship 1XWM report in the Frontier Labs section above.
Production deployment case studies
5Companies talking about policies running in actual warehouses and facilities, not lab demos — the closest thing available to ground truth on what "100% reliable" looks like when the customer is watching.
RFM-1 (product page)
Trained on "tens of millions" of trajectories from live warehouse fleets, named customers include Crate & Barrel and Bonprix. Follow-up on in-context learning for grasping.
An OpenAI Spinoff Has Built an AI Model That Helps Robots Learn Tasks Like Humans
Independent press coverage of Covariant RFM-1 — useful as an outside-in read on the same result reported by the company itself above.
Ambi Robotics Deploys PRIME-1
Unified transformer backbone for 3D perception, picking, and QC, fine-tuned on 4 years of proprietary warehouse data. Claims fleet-wide collective learning improves reliability over time, not just per-unit accuracy.
AgiBot Releases Humanoid Manipulation Dataset
Trade coverage of the AgiBot World release, listed here for the deployment/business framing rather than the technical detail (which is under Datasets above).
Dexterity blog
Reports 200+ unique items picked in production at 99.5% accuracy, 50,000+ SKUs handled, 14M+ picks performed — including FedEx bulk-truck-loading. Force/touch-controlled, in production, not academic — worth crawling directly for more.
Reflections, surveys & reliability research
10People who've built enough of these systems to say plainly what isn't working yet, plus the small but growing thread of work that treats long-tail reliability as its own research problem rather than a side effect of scale.
The Promise of Generalist Robotic Policies
Reflection from one of the field's core RL/IL researchers on what generalist VLA-style policies actually promise versus what they've delivered so far.
Robot Learning: A Tutorial
Broad survey/tutorial covering the field's current methods end to end — a good on-ramp if you want the textbook version alongside these primary sources.
Robot Learning in the Era of Foundation Models: A Survey
Peer-reviewed survey positioning current VLA/foundation-model work within the longer arc of robot learning research.
The Data Gap That's Holding Back Robotics
Outside-in explainer on the same data bottleneck Scale AI names directly above — useful as a plain-language framing to hand to someone less deep in the field.
RaC: Scaling Recovery and Correction for Long-Horizon Tasks
Directly targets the "demos show success paths, not recovery from failure" gap by deliberately scaling recovery/correction data rather than only success data.
State of Robot Learning, December 2025
The single most useful practitioner overview in this collection. Candid claim: robot learning today is still overwhelmingly behavior cloning. Walks through the three real data-collection paradigms (leader-follower/ALOHA-style, wearable interfaces like UMI, human video), explains why DAgger-style recovery-data collection matters, and why RL's LLM-era success hasn't transferred to robotics — you can't cheaply reset a cluttered real-world scene the way you can reset a game environment.
Deployment-Time Reliability of Learned Robot Policies
Directly on-topic for the "100%" question: argues reliability at deployment isn't a property of model class or data scale alone, but emerges from reliability-promoting mechanisms spread across the whole autonomy stack and development lifecycle.
ROBOGATE: Adaptive Failure Discovery for Safe Robot Policy Deployment
Two-stage boundary-focused sampling to actively find a policy's failure modes before deployment, rather than discovering them in production.
This Is the State of the Art in Robotics
Talk covering where the field's ceiling actually sits right now, from one of the researchers behind ALOHA/Mobile ALOHA/Diffusion Policy-adjacent work.
FAR: Failure-Aware Retry for Test-Time Recovery
Test-time retry/recovery mechanism — treats reliability as something you can improve at inference time, not only through more training data.