🤖 Another zero-shot reward model is now in LeRobot:... ROBOMETER. A general-purpose, zero-shot video-language reward model from University of South Carolina, UT Dallas, Massachusetts Institute of Technology (MIT), University of Washington, Ai2, and NVIDIA that predicts frame-level task progress. Trained on 1M+ trajectories from 21 robot embodiments, generalizes zero-shot to unseen tasks, scenes, and robots. 2.4–4.5x better downstream success rates across online RL, offline RL, data filtering, failure detection, and data retrieval for IL. Project: Paper:show more

Robotics keeps hitting the same wall. Single task RL... works, but... it does not scale to hundreds of tasks or new embodiments. This new paper looks like a real step toward fixing that. The team introduces MMBench, a benchmark with 200 tasks across many domains and robots, and Newt, a language conditioned world model trained online across all 200 tasks at once. The simple idea behind Newt: The model learns from demos to get the right priors It trains across many tasks through online interaction It uses language to ground the goal It adapts fast when a new task shows up What stood out to me: ✅ One model trained on 200 tasks at the same time ✅ Language conditioned control for both states and RGB ✅ Better data efficiency than strong baselines ✅ Strong open loop control ✅ Fast adaptation to new tasks and embodiments ✅ Full release of 200 checkpoints, 4000 demos, code, and benchmark This is a good push toward general control instead of one model per task. If you want the full paper: Project page: —- Weekly robotics and AI insights. Subscribe free:show more

🤖 NVIDIA’s Gr00t N1.5 is now available in LeRobot!... This is the result of a great collaboration between the Hugging Face LeRobot team and NVIDIA Robotics ! Gr00t N1.5 highlights: 🦾 Cross-embodiment foundation model for robots 🧠 Multimodal inputs: vision, language, and proprioception 🪛Tested on the Libero benchmark and real-world hardware tasks 🌍Trained on real robot, synthetic, and internet-scale video data ⚙️ Flow matching action transformer for action predictionshow more

Robot Utility Models (RUMs) enable basic tasks – door... opening, drawer opening, object reorientation, etc. – at ~90% accuracy without ANY finetuning (i.e. zero-shot) in unseen new environments. Fully open source!!! models, data, code & hw. We think this is super exciting, why?👇 1. Unlocks many practical home utility tasks that often involve these basic tasks as part of an action chain. “Go get me a fork” involves opening the kitchen door and then opening the cutlery drawer. 2. This works well **zero-shot in unseen and new** environments, which is practically a huge deal. Turn the robot on, and get going. 3. The recipe for building a new model is fairly generic, and we think with a bit more refinement this can be a general recipe to build many more Utility models. More details and access 👇show more

30 minutes of video. Robot learns the task. Open-source,... end-to-end. An open-source framework for training robot policies from only 30 minutes of human egocentric videos captured via Meta Aria glasses: Achieving zero-shot transfer to robots without any robot data collection. The method relies on Interaction-Centric Tokens that encode hand-object spatial relationships invariant to embodiment and viewpoint, supplemented by auxiliary objectives like object motion prediction and latent consistency to extract richer supervision signals from the same data. HumanEgo demonstrates strong cross-embodiment, cross-environment performance on bimanual tasks, outperforming baselines like ACT and teleop data while being trainable on a single RTX 4090 GPU. Thanks for sharing, Zhi (Leo) Wang. 📌 Website: Paper: Code: Video: ——- Weekly robotics and AI insights. Subscribe free:show more

Trained on zero real-world data. Learned to walk, pick... up boxes, and follow multi-step instructions... in the REAL world. ( 📌 Paper below) Researchers from Amazon FAR, Berkeley, Stanford, and CMU scanned real rooms with an iPhone, rebuilt them as 3D Gaussian Splatting scenes, then generated 48,000 synthetic trajectories of a Unitree G1 walking, grasping, and placing objects inside those virtual replicas. They rendered the robot's first-person camera view from each run and paired it with the matching language instruction and motion data. That's the dataset every humanoid team needs and nobody has: synced egocentric video + language + kinematics, at scale. Instead of collecting it in the real world, they manufactured it. They trained a vision-language-kinematics policy on that synthetic data alone, then deployed it on the physical G1 across five task types: navigation to a named object, lifting boxes of three different sizes with no per-size tuning, chained multi-step tasks, robustness to mid-task layout changes and flickering lights, and multi-minute long-horizon runs. No real-world fine-tuning at any point. Real-world interaction data has been the hard limit on humanoid learning... slow, expensive, and small. If scanning a room once and synthesizing thousands of labeled interactions holds up as a general recipe, that limit moves. Data stops being the bottleneck robotics teams have to solve for. 📌 Paper: Project: ——- Weekly robotics and AI insights. Subscribe free:show more

Most imitation learning policies break when the camera moves... or the robot changes. NOT THIS ONE 👇 [📍 Bookmark for later ] A new 3D scene representation encoder, tackles this by enabling zero-shot generalization to unseen embodiments and viewpoints… And it works with any IL algorithm. The trick? •Use a 2D foundation model to extract semantic features •Lift them into 3D space for localization (not semantics) •Condition the IL policy on this spatially grounded vector Across 93 simulated and 6 real tasks, Adapt3R: ✅ Maintains IL performance on LIBERO & MimicGen benchmarks ✅ Outperforms DP3 and 3D Diffuser Actor in most settings ✅ Holds >80% success on LIBERO even with large camera rotations Thanks for sharing this, Animesh Garg & Albert Wilcox! 📍Paper: Website: Code:show more

Fuck yeah! MaskGCT - New open SoTA Text to... Speech model! 🔥 > Zero-shot voice cloning > Emotional TTS > Trained on 100K hours of data > Long form synthesis > Variable speed synthesis > Bilingual - Chinese & English > Available on Hugging Face Fully non-autoregressive architecture: > Stage 1: Predicts semantic tokens from text, using tokens extracted from a speech self-supervised learning (SSL) model > Stage 2: Predicts acoustic tokens conditioned on the semantic tokens. Synthesised: "Would you guys personally like to have a fake fireplace, an electric one, in your house? Or would you rather have a real fireplace? Let me know down below. Okay everybody, that's all for today's video and I hope you guys learned a bunch of furniture vocabulary!" TTS scene keeps getting lit! 🐐show more

Introducing VL-JEPA: Vision-Language Joint Embedding Predictive Architecture for streaming,... live action recognition, retrieval, VQA, and classification tasks with better performance and higher efficiency than large VLMs. • VL-JEPA is the first non-generative model that can perform general-domain vision-language tasks in real-time, built on a joint embedding predictive architecture. • We demonstrate in controlled experiments that VL-JEPA, trained with latent space embedding prediction, outperforms VLMs that rely on data space token prediction. • We show that VL-JEPA delivers significant efficiency gains over VLMs for online video streaming applications, thanks to its non-autoregressive design and native support for selective decoding. • We highlight that our VL-JEPA model, with an unified model architecture, can effectively handle a wide range of classification, retrieval, and VQA tasks at the same time. by Delong Chen (陈德龙) Mustafa Shukor Théo Moutakanni Willy Jade Lei Yu Tejaswi Kasarla Allen Bolourchi Yann LeCun Pascale Fungshow more

If you’ve been ignoring Axis Robotics because it looks... like another random points farm, read this. In the last 48 hours, Unitree said robotics is approaching its "ChatGPT moment", while the chairman of ACE Robotics believes it could happen by the end of 2027. But for robots to reach that level, they need a crazy amount of training data. That’s exactly what Axis is building. You control simulated robots directly from your browser, complete simple tasks and earn points. Those movements also help create data that can be used to train real robots. And this isn’t some tiny experiment anymore: - $12M raised - 123K+ contributors - 3M+ robot trajectories - Community data already used to train a real robot So yeah, we’re basically farming a potential airdrop while teaching our future robot servants how to work 😂 If you haven’t started yet, it’s completely free. You only need a tiny amount of gas on Base to sign your completed tasks. ✅ Start farming Axis points: Important: Sign every completed task from the History page, otherwise you won’t receive the points.show more

🔥 Nebius AI R&D is hiring AI Research Interns... for short, high-impact RL projects. Exclusive to X right now — no LinkedIn mass postings yet. In 2019, I was a fresh dental grad with 3 months of runway left, begging for an AI shot. I know the grind. We’re looking for sharp early-career folks (students, grads, career-switchers) to join us and work on: > Agent trajectories analysis at scale > Long-horizon tasks for coding agents > Pushing open RL environments > Any other data / RL env / eval project that will benefit open-source community What you get: 💰 Fully paid internship (3-6 month) 📦 100% open-source shipping 📄 Co-author research papers ⚡️ Access to Nebius compute infra 🌍 Remote-friendly (EU/US) or Amsterdam/London/other office. If you’ve done any cool AI/ML/RL stuff, dm me with your most impressive project + 1-sentence summary + cv Sharing appreciated!🤝show more

𝗥𝗼𝗯𝗼𝘁𝘀 𝗱𝗼𝗻’𝘁 𝗻𝗲𝗲𝗱 𝗺𝗼𝗿𝗲 𝗱𝗲𝗺𝗼𝗻𝘀𝘁𝗿𝗮𝘁𝗶𝗼𝗻𝘀. 𝗧𝗵𝗲𝘆 𝗻𝗲𝗲𝗱 𝘁𝗼 𝗹𝗲𝗮𝗿𝗻... 𝗳𝗿𝗼𝗺 𝗳𝗮𝗶𝗹𝘂𝗿𝗲 — 𝗮𝗳𝘁𝗲𝗿 𝘄𝗮𝘁𝗰𝗵𝗶𝗻𝗴 𝗵𝘂𝗺𝗮𝗻𝘀. Most robot learning systems assume failure is the end of learning. In our new work, we study whether robots can improve after deployment by learning from their own failures, without any human intervention, teleoperation, or corrective labels. The key idea is simple: human videos contain structure about how the world works. We use them to learn cross-embodiment representations of action, dynamics, and value, enabling a shared predictive space between human behavior and robot experience. This allows a new learning loop: 👉 pretrain on human videos 👉 deploy robot policy 👉 observe failures 👉 reinterpret failures using human priors 👉 improve autonomously We evaluate this across 7 real-world manipulation tasks, showing: 📈 40% → 81% success rate 🏆 Strong improvements over π0.6 RECAP and RISE ✔️ Zero human intervention during post-deployment improvement 🧬 Generalizes across robot embodiments and policy backbones A key finding is that explicit failure repair significantly outperforms failure reweighting, yielding substantially larger gains under identical data conditions (+25 pts vs +5 pts on the same π0.5 base policy). Overall, the results suggest a shift in how we think about robot learning: Human videos are not only for pretraining policies. They can provide the structure needed for continual self-improvement after deployment. 📄 Paper: 🌐 Project: I am grateful for working with the fantastic leads Hanzhi Chen and Anran Zhang, and our collaborators Simon Schaefer, Kejia Chen, Shi Chen, Daniel Cremers. Special thanks to Stefan Leutenegger for co-advising this project with me. ETH Zürich TU München Microsoft Check out Hanzhi's 🧵 for more detailsshow more

[Most robots react. This one thinks a step ahead.]... Ant Group's Robbyant just published LingBot-VA 2.0 — a video-action foundation model built from scratch for robot control, not fine-tuned from a video generator. The usual approach takes a video generator made for content creation and bolts a robot policy onto it. LingBot-VA 2.0 argues that's the wrong starting point, and pretrains the whole causal stack natively instead. What stands out: → Foresight Reasoning — the robot predicts the next action chunk while executing the current one, then overwrites the imagined frame with the real observation. Prediction and execution stop waiting on each other. → 927 ms → 142 ms per chunk, across four cumulative optimizations. That lifts asynchronous control from 35 Hz to 225 Hz — a 6.5× speedup. → One shared latent space. A semantic visual-action tokenizer puts world states and actions in the same coordinates, so unlabeled web video carries action-relevant signal. → Sparse MoE video stream — 128 experts, top-8 routing. Roughly 2.5B of ~15.3B parameters fire per token. → Few-shot by design — adapts from 10–15 demonstrations, and a human demo video can replace the text instruction entirely. Full breakdown: Paper: Project Page: Robbyant Ant Groupshow more

Yes they did. Btw here is what a .30-06... does do a resin skull encased in ballistic gel. Charlie Kirk was murdered by professional hitmen and the FBI blamed a patsy. Candace please discuss some of these inconsistencies. Amateur Shooter Skill Discrepancy: Official narrative claims Tyler Robinson, with no professional training, executed a precise 142-yard neck shot using a vintage Mauser 98 rifle. Ballistics data shows a miniscule success rate for novices. He had zero practice records or ammo traces. single-shot perfection is impossible without expert help. Invisible Rooftop Presence and Firing: No eyewitnesses saw Robinson climbing, positioning, or shooting despite ~3,000 attendees and visible rooftop in daylight. Witness statements point to conflicting shot directions. The probability of zero sightings 2 minutes Video shows fleeing figure, but actions exceed feasible time for amateur. Post-Escape Behavior: Calm demeanor post-shot he was at Dairy Queen posting Discord memes. Only arrested after dad's tip-off. His "Confession" text to vanished lover, unreleased geo-tracking data, amateur would show panic, not casual behavior. Audio Discrepancy across Multiple Sources: Videos show varying shot sounds (immediate crack in close views, delays/echoes in others). Physics predicts consistent crack-bang from a single rooftop. Variations imply closer/suppressed weapon or multiple shooters. Trajectory and Wound Mismatches: Blood spray arcs forward/left, not downward from elevated shot. No exit despite high velocity. Some Witnesses report "two shots." Evidence Tampering and Scene Handling: Stage/cameras dismantled ~5-10 mins post-shot. FBI delayed ~16 mins and no autopsy released. Videos show staff yanking SD cards, there are gag orders on witnesses and the scene was trampled by people who were supposed to be professionals. Pre-Event Foreknowledge Indicators: Google searches for "Tyler Robinson" spiked pre-shot. Hand signals in videos, scrubbed CIA-linked professors. Suggests coordination/setup.show more

Human3R: Everyone Everywhere All at Once Note: I recorded... the video from the interactive demo on their project page (linked in the comment below). Abstract (excerpt): Human3R jointly recovers global multi-person SMPL-X bodies ("everyone"), dense 3D scenes ("everywhere"), and camera trajectories in a single forward pass ("all-at-once"). Our method builds upon the 4D online reconstruction model CUT3R and uses parameter-efficient visual prompt tuning to preserve CUT3R's rich spatiotemporal priors while enabling direct readout of multiple SMPL-X bodies. Human3R is a unified model that eliminates heavy dependencies and iterative refinement. After being trained on the relatively small-scale synthetic dataset BEDLAM for just one day on one GPU, it achieves superior performance with remarkable efficiency: it reconstructs multiple humans in a one-shot manner, along with 3D scenes, in one stage, at real-time speed (15 FPS) with a low memory footprint (8 GB).show more

Furniture assembly is the task everyone name-drops and nobody... actually attempts at real scale. Every demo I have seen is a scaled down IKEA leg or a single arm on a toy chair. This paper does it properly, real scale, bimanual, up to 7 subtasks and 1,550 control steps per episode, and it is validated on a real Kinova Gen3, not just in sim. That real-robot number is the one that matters: only a 16 percent drop on the hardest task going from simulation to hardware. That is a small enough gap to take seriously, and it did not happen by accident. They built a VR teleoperation rig specifically for coordinated dual-arm collection, because generic single-arm teleop setups do not capture the coordination real assembly needs, and the model predicts a continuous progress signal alongside the action chunk rather than a discrete subtask label, letting it auto-transition and catch drift before it compounds into total failure. The simulation ablation is what got them there, 48 to 80 percent over baselines, with another 21 points from their perception and control design study alone, but that is groundwork, not the headline. Watch the video, there is a clip of the robot misgrasping the seat panel, reopening the gripper, and regrasping on its own. That is not scripted recovery behaviour, it emerged from training, and it emerged on hardware. Excellent work from the team from Mitsubishi Electric Research Laboratories, with Oxford and UNC Chapel Hill Clinical Laboratory Science. Video and project page in comments. #Robotics #Manipulation #VLAshow more

🛠️ What if a robot could invent its own... tools. And teach itself how to use them? That’s exactly what VLMgineer does: a new framework that lets Vision Language Models (VLMs) design physical tools and the actions to use them, entirely on their own. No templates. No human demonstrations. Just raw, AI-driven creativity. Why it matters ✅ Co-designs tools and actions together using VLMs, ensuring tight coupling between form and function ✅ Uses VLM-guided evolution (not random search) to refine designs intelligently ✅ Outperforms human-designed tools by +64.7% in task success across 12 RoboToolBench challenges ✅ Produces better-than-everyday tools for real manipulation tasks—measured in success rate and elegance It builds on the emerging trend of large-model-guided evolutionary design (like Eureka and AlphaEvolve) and brings it into physical robotics. It opens the door to general-purpose, automated hardware design, no strong priors needed. Code & paper: —- Weekly robotics and AI insights. Subscribe free:show more

𝗗𝗼𝗻'𝘁 𝗳𝗶𝗻𝗲-𝘁𝘂𝗻𝗲 𝗿𝗼𝗯𝗼𝘁 𝗳𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻 𝗺𝗼𝗱𝗲𝗹𝘀. 𝗦𝘁𝗲𝗲𝗿 𝘁𝗵𝗲𝗺 𝘄𝗶𝘁𝗵 𝗵𝘂𝗺𝗮𝗻... 𝗰𝗼𝗿𝗿𝗲𝗰𝘁𝗶𝗼𝗻𝘀 𝗶𝗻𝘀𝘁𝗲𝗮𝗱, 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗰𝗵𝗮𝗻𝗴𝗶𝗻𝗴 𝘁𝗵𝗲 𝗯𝗮𝘀𝗲 𝗽𝗼𝗹𝗶𝗰𝘆 Modern VLAs and world-action models can perform impressive manipulation skills, but adapting them reliably to new robots and tasks remains challenging. A natural solution is DAgger-style online imitation learning: deploy the robot, collect human corrections, and update the policy. Yet foundation models are fragile in the low-data regime, fine-tuning on a handful of interventions can improve one behavior while degrading others. Online post-training or reinforcement learning can require costly data collection and exploration, making real-world learning expensive and potentially unsafe. In our new paper, 𝗙𝗹𝗼𝘄𝗗𝗔𝗴𝗴𝗲𝗿, we take a different approach: 𝗜𝗻𝘀𝘁𝗲𝗮𝗱 𝗼𝗳 𝗰𝗵𝗮𝗻𝗴𝗶𝗻𝗴 𝘁𝗵𝗲 𝗳𝗼𝘂𝗻𝗱𝗮𝘁𝗶𝗼𝗻 𝗺𝗼𝗱𝗲𝗹, 𝘄𝗲 𝗹𝗲𝗮𝗿𝗻 𝗵𝗼𝘄 𝘁𝗼 𝘀𝘁𝗲𝗲𝗿 𝗶𝘁 𝗳𝗿𝗼𝗺 𝗵𝘂𝗺𝗮𝗻 𝗰𝗼𝗿𝗿𝗲𝗰𝘁𝗶𝗼𝗻𝘀. The key idea is 𝗮𝗰𝘁𝗶𝗼𝗻 𝗶𝗻𝘃𝗲𝗿𝘀𝗶𝗼𝗻: we map human corrective actions back into the latent noise space of the frozen generative policy. These latent targets train a lightweight controller that adapts the robot while preserving the original model's capabilities. Across simulation and real robots, FlowDAgger: 📈 Learns from only 5–20 human intervention episodes 🏆 Outperforms supervised fine-tuning and latent-space reinforcement learning 🤖 Works across VLAs, diffusion policies, and world-action models ✔️ Provides reliable improvements without modifying the pretrained policy We believe this offers a practical path toward making robot foundation models improve during deployment, learning from the way humans naturally teach: through corrections. 📄 Paper: 🌐 Project: 💻 Code: This project was led by my amazing colleague Michael Murray with help from Daphne Chen, Simran Bagaria, Dean Fortier, Tess Hellebrekers, Harshavardhan Reddy Gajarla, Galen Mullins and Andrey Kolobov at Microsoft Research and Maya Cakmak at University of Washingtonshow more

Wonderland: Navigating 3D Scenes from a Single Image Contributions:... • First, we introduce a representation for controllable 3D generation by leveraging the generative priors from camera-guided video diffusion models. Unlike image models, video diffusion models are trained on extensive video datasets. This enables them to capture comprehensive spatial relationships within scenes across multiple views and embed a form of "3D awareness" in their latent space, which allows us to maintain 3D consistency in novel view synthesis. • Second, to achieve controllable novel view generation, we empower video models with precise control over specified camera motions. We introduce a novel dual-branch conditioning mechanism that effectively incorporates desired diverse camera trajectories into the video diffusion model. This enables expansion of a single image into a multi-view consistent capture of a 3D scene with precise pose control. • Third, to achieve efficient 3D reconstruction, we directly transform video latents into 3DGS. We propose a novel latent-based large reconstruction model (LaLRM) that lifts video latents to 3D in a feed-forward manner. With this design, during inference, our model directly predicts 3DGS from a single input image, effectively aligning the generation and reconstruction tasks—and bridging image space and 3D space—through the video latent space. Compared with reconstructing scenes from images, the video latent space offers a 256× spatial-temporal reduction while retaining essential and consistent 3D structural details. Such a high degree of compression is crucial, as it allows the LaLRM to handle a wider range of 3D scenes within the reconstruction framework, with the same memory constraints.show more

🚀 Introducing EgoExo Forge - built on top of... Rerun, Gradio, and Hugging Face hub (I’ll be in San Francisco July 21–29 — if you’re into robotics, egocentric AI, large-scale data collection, or just want to chat, DM me!) In my opinion, large-scale, diverse, and high-quality data is still the largest bottleneck for generalized robotics deployment. I believe that some version of imitation learning from human examples will be the most scalable + clean way to train humanoid robots 🤖 (similar to what Tesla did for Full Self Driving). Teleop is too expensive to collect a large enough dataset in a reasonable manner, so passive collection via egocentric (and in certain cases, exocentric) views feels like the right bet. Over the past few months, I've been trying to build out the scaffolding for this and using Rerun as my underlying infrastructure. Data being collected needs to be easily inspectable + time series and rerun provides the right tooling for this. My goal is to first build out a ground truth representative dataset from already existing open source data, generate some reasonable baselines, and then go out and collect my own data that adheres to the defined schema. 🔍 Starting with open-source datasets 1. EgoDex from Apple 2. HOCap from Nvidia and the University of Texas at Dallas 3. Assembly101 from Meta All these different datasets have different sensor configurations + annotations, so my goal with egoexo-forge is to have one consistent labeling scheme + data layout. I built a data pipeline that aligns all of the different datasets in one general schema assuming the COCO133 keypoint layout that allows for exo+ego, ego only, or exo only Since the scaffolding is already there, it becomes MUCH easier to add other datasets. So the next ones that I'll be including are HD-EPIC kitchens dataset, HOT3D, and finally my own personal iPhone + insta360 go collection method. Once I have a diverse variety of datasets, I'll double down on what I believe to be the key algorithms required to make useful data for imitation learning 📊 1. Camera Pose estimation via SLAM/SFM for ego perspective (and automatic calibration for exo) 2. Human pose estimation for both egocentric + exocentric views 3. Metric 3D reconstruction + object tracking I'll be setting up reasonable open-source baselines for each of these to validate that these datasets work, and then finally try to use the generated datasets for some imitation learning via the pi0-lerobot repo I've been working on. I plan on making a blog post + providing more info on all of this in the near future so stay tunedshow more

Model-Free Reinforcement Learning (MFRL) has been alluring, especially with... supercharged compute with physics on GPU. However, the methods use 0-th order gradients, and are often not the best optimizers. Can we do better than PPO in continuous control for robotics? Turns out yes! 🥳 tl;dr: Faster, better RL than PPO in continuous control 💪 The answer lies in using more information from the simulation. We are juicing the simulation on GPU as it is, why not use it for gradients as well? This has been a driving question in a series of our works. We first studied this problem in ICLR 2022 paper on Short Horizon Actor Critic Naive gradient based methods are stuck in local minima and have exploding/vanishing gradients. SHAC solved this problem truncated rollouts and model based value estimation, where the model is Differentiable Sim. This boosted sample efficiency and wall-clock time immensely especially in high dimensional systems such as humanoids Yet, given enough compute PPO often caught up. Our follow up paper on on Adaptive Horizon Actor Critic at ICML 2024 discovers the cause and provides a fix. However, we find that even when given ground-truth dynamics, not all gradients are useful due to sample error. 1st-Order Model-Based Reinforcement Learning methods employing differentiable simulation provide gradients with reduced variance but are susceptible to bias in scenarios involving stiff dynamics, such as physical contact. We find that back-propagating through contact and long trajectories drastically reduces gradient accuracy. Using this insight, we propose AHAC to dynamically adapt its roll-out horizon to avoid differentiating through stiff contact. AHAC is a first-order model-based RL algorithm that learns high-dimensional tasks in minutes (wall clock) and outperforms PPO by 40%, even in the limit of data provided to PPO. This work is led by Ignat Georgiev alongside Krishnan Srinivasan, Jie Xu, Eric Heiden and ample assistance from warp team at NVIDIA Robotics (Miles Macklin)show more