https://frontieraianalysis.com/video/a-video-world-model-can-remember-through-changing-weights/https://frontieraianalysis.com/assets/images/editorial-covers-v2/photo-000-77bcf95cde.jpgWorld-T3 turns changing weights into visual memoryThe World-T3 team’s ten-second living-room revisit rollout. This is generated model output, not footage of a physical experiment; the project page offers the reference video and competing methods for comparison.https://adreamwu.github.io/World-T3/static/videos/revisit_cmp/case1/ours.mp410https://frontieraianalysis.com/video/robot-memories-need-an-address-not-just-a-description/https://frontieraianalysis.com/assets/images/cover-upgrade/real-001-7acea01d26.jpgRobot memories need an address, not just a description — research demonstrationMemoryVLN combines what a robot saw with where it stood. Official project demonstration; the selected clip is illustrative footage, not an independent benchmark.https://zgca-hmi-lab.github.io/MemVLN/static/videos/real_task1.mp4178https://frontieraianalysis.com/video/a-robot-s-memory-should-help-it-decide-what-to-look-at-next/https://frontieraianalysis.com/assets/images/cover-upgrade/real-002-be3274012e.jpgA robot's memory should help it decide what to look at next — research demonstrationMemAct puts retrieval and exploration in the same loop. Official project demonstration; the selected clip is illustrative footage, not an independent benchmark.https://vfishc.github.io/memact/assets/overview.mp4181https://frontieraianalysis.com/video/more-ai-researchers-do-not-automatically-make-better-research/https://frontieraianalysis.com/assets/images/editorial-covers-v2/007-more-ai-researchers-do-not-automatically-make-better-research.jpgAreal: RL Infra for OpenClawBackground: Wei Fu discusses AReaL’s reinforcement-learning infrastructure in a NICE AI Talk. The lecture explains the underlying training system; it is not a demonstration or validation of the later AReaL-RSI multi-agent results.https://www.youtube-nocookie.com/embed/C-1po0fKbNAhttps://frontieraianalysis.com/video/a-long-agent-task-can-be-made-learnable-by-exposing-its-hidden-progress/https://frontieraianalysis.com/assets/images/editorial-covers-v2/008-a-long-agent-task-can-be-made-learnable-by-exposing-its-hidden-progres.jpgA long agent task can be made learnable by exposing its hidden progress — research demonstrationTev-Gym builds verifiable environments from an explicit latent specification. Official project demonstration; the selected clip is illustrative footage, not an independent benchmark.https://tev-agent.github.io/videos/tev_launch_v6.mp480https://frontieraianalysis.com/video/short-term-computation-and-long-term-memory-can-use-different-machinery/https://frontieraianalysis.com/assets/images/editorial-covers-v2/025-short-term-computation-and-long-term-memory-can-use-different-machiner.jpgSakana separates short-term computation from long-term memoryBackground: Sakana AI’s 2025 animation explains the Continuous Thought Machine, the architecture on which the later Continuous Memory Machine builds. It does not demonstrate or validate CMM’s new long-term-memory results.https://pub.sakana.ai/ctm/assets/mp4/blog_arch_video.mp459https://frontieraianalysis.com/video/improving-the-answer-and-improving-the-next-workflow-are-different-loops/https://frontieraianalysis.com/assets/images/editorial-covers-v2/064-improving-the-answer-and-improving-the-next-workflow-are-different-loo.jpgImproving the answer and improving the next workflow are different loopsThe Harness² authors’ mechanism animation, published in the official repository’s video overview. It illustrates the framework and paper-reported results; it is not a recording of a new independent experiment.https://github.com/user-attachments/assets/76b10548-b14b-4c8f-9aa5-a93662df4d5647https://frontieraianalysis.com/video/gemini-argon-apollo-video-builds-an-experiment/https://frontieraianalysis.com/assets/hover-gifs/gemini-argon-apollo-video-builds-an-experiment-cornhole.jpgTeaching a humanoid robot to play cornhole with Gemini 4 ArgonGoogle’s official demonstration of Apptronik’s Apollo humanoid throwing cornhole beanbags after Gemini 4 Argon builds a simulation from human video. The reported seven consecutive successful throws are Google’s internal result.https://www.youtube-nocookie.com/embed/9PICEmtfpWU2026-10-13103https://frontieraianalysis.com/video/nvidia-v2d-reconstruction-errors-robot-learning/https://frontieraianalysis.com/assets/hover-gifs/nvidia-v2d-reconstruction-errors-robot-learning.jpgVideo to Data Challenge: From Human Video to Robot SkillsThe video linked from NVIDIA’s Video-to-Data challenge page explains the challenge and its workflow.https://www.youtube-nocookie.com/embed/U8GHIxXqHLQ2026-10-077410https://frontieraianalysis.com/video/liquid-d1-the-cost-of-reading-the-world/https://frontieraianalysis.com/assets/hover-gifs/liquid-d1-the-cost-of-reading-the-world.jpgLiquid AI can answer more questions without looking at the scene againLiquid AI’s published d1 demonstration. The recording is shown at 4× speed; it is not a real-time latency measurement.https://aypchzzf9pftwuto.public.blob.vercel-storage.com/Screen%20Recording%202026-10-06%20at%2011.11.14%E2%80%AFPM_4x-2AJepHDZSIDeceCSmCMkWcCsA4VBBA.mp42026-10-0742https://frontieraianalysis.com/video/physical-intelligence-api-next-demonstration/https://frontieraianalysis.com/assets/hover-gifs/physical-intelligence-api-next-demonstration.jpgPhysical Intelligence makes robot AI easier to access. What should you teach it next?Almond’s Axol robot deburrs spacers used in its own assembly. Physical Intelligence presents this demonstration at 1× speed.https://website.pi-asset.com/api/almond-action.mp42026-10-13179https://frontieraianalysis.com/video/egomatrix-contact-before-photorealism/https://frontieraianalysis.com/assets/hover-gifs/egomatrix-contact-before-photorealism.jpgRobot training needs realistic contact, not just realistic picturesA real-robot flower-handling demonstration from the EgoMatrix project.https://spikew726.github.io/EgoMatrix/assets/real-v2/flowers.mp435https://frontieraianalysis.com/video/augustus-driving-reasoning-without-narration/https://frontieraianalysis.com/assets/images/editorial-covers-v2/older-106-augustus-driving-reasoning-without-narration.jpgA driving AI can learn from explanations without talking through every moveThe researchers’ first real-vehicle driving demonstration, published alongside AUGUstus. This selected drive is illustrative, rather than an independent safety evaluation.https://raw.githubusercontent.com/haha-yuki-haha/AUGUstus/main/assets/b19.mp43https://frontieraianalysis.com/video/inspect-robots-reset-is-part-of-the-test/https://frontieraianalysis.com/assets/hover-gifs/inspect-robots-reset-is-part-of-the-test.jpgHow you set up a robot test can change its resultA pouring-task comparison from RoboHarm, one of the evaluation suites discussed in Inspect Robots. This clip illustrates that suite, rather than the Inspect Robots software itself.https://robocurve.org/roboharm/video/task-pour-grid.mp4?v=07af5487a4fa10https://frontieraianalysis.com/video/humanoidtoolbench-right-tool-wrong-body/https://frontieraianalysis.com/assets/hover-gifs/humanoidtoolbench-right-tool-wrong-body.jpgA robot must be able to handle a tool, not just choose itThe project’s overview demonstrates humanoid tool selection and execution.https://snu-pi.github.io/HumanoidToolBench/static/video/humanoidtoolbench.mp4179https://frontieraianalysis.com/video/openvitac-test-whether-touch-changes-the-action/https://frontieraianalysis.com/assets/hover-gifs/openvitac-test-whether-touch-changes-the-action.jpgA robot with touch sensors may still rely on sightA successful real-world USB-insertion trial from OpenViTac. The project also publishes failure examples; this selected clip is not an aggregate performance measure.https://fvl-repo.github.io/OpenViTac/assets/demos/real/insert_usb/standard/success/demo.mp421https://frontieraianalysis.com/video/sota-sensor-design-is-a-data-format/https://frontieraianalysis.com/assets/hover-gifs/sota-sensor-design-is-a-data-format.jpgMatching touch sensors could help robots learn from human handsSoTa’s cup-handling demonstration using its co-trained policy. See the project for the baseline comparison.https://sota-skin.github.io/assets/cup-cotrain.mp412https://frontieraianalysis.com/video/evo-rl-successful-trajectories-bad-lessons/https://frontieraianalysis.com/assets/hover-gifs/evo-rl-successful-trajectories-bad-lessons.jpgA successful robot run can still teach the wrong lessonA policy rollout animation from the official Evo-RL repository, converted from GIF to video for playback. It illustrates the released pipeline; it is not the paper’s full comparison of selective experience.https://frontieraianalysis.com/assets/article-videos/evo-rl-policy-rollout.mp440https://frontieraianalysis.com/video/fastopd-compression-changes-the-path/https://frontieraianalysis.com/assets/hover-gifs/fastopd-compression-changes-the-path.jpgFastOPD: Making robot intelligence smaller without losing the actionThe researchers compare the teacher and student policies in real-robot trials.https://fastopd.github.io/static/videos/sample1.mp412https://frontieraianalysis.com/video/ttt-rm-memory-for-what-the-camera-cannot-see/https://frontieraianalysis.com/assets/hover-gifs/ttt-rm-memory-for-what-the-camera-cannot-see.jpgRobot memory is most useful when the camera cannot see the answerThe project’s supplementary video illustrates its robot memory experiments.https://hatchetproject.github.io/tttrm/static/videos/supplementary.mp492https://frontieraianalysis.com/video/tapnav-touch-as-a-navigation-experiment/https://frontieraianalysis.com/assets/hover-gifs/tapnav-touch-as-a-navigation-experiment.jpgA longer route can help a robot find its way by touchTapNav’s project video shows tactile navigation and the experimental setup.https://tapnav.github.io/TAPNAV-Project/letsgo.mp4180https://frontieraianalysis.com/video/openwam-less-context-better-transfer/https://frontieraianalysis.com/assets/hover-gifs/openwam-less-context-better-transfer.jpgGiving part of a robot's AI less information helped it learn a new taskThe OpenWAM team demonstrates bread toasting and completing the final layer of a small Rubik’s cube.https://openwam.stanford.edu/assets/openwam-robot-teaser.mp435https://frontieraianalysis.com/video/tacross-cheap-touch-data-needs-a-translator/https://frontieraianalysis.com/assets/hover-gifs/tacross-cheap-touch-data-needs-a-translator.jpgA cheap sensor glove still needs robot demonstrations to teach actionsSupplementary demonstrations of the TACROSS tactile data collection and robot tasks.https://tacross-touch-project.github.io/assets/supplementary-demonstrations.mp4179https://frontieraianalysis.com/video/embeddinggemma2-shared-space-model-building/https://frontieraianalysis.com/assets/hover-gifs/embeddinggemma2-shared-space-model-building.jpgEmbeddingGemma 2: Better answers start with finding the right evidenceGoogle’s official AI Edge Gallery demonstration of media retrieval with EmbeddingGemma 2.https://storage.googleapis.com/gweb-uniblog-publish-prod/original_videos/google-ai-edge-gallery-instant-media-search_eRLOpvE.mp42026-10-0629https://frontieraianalysis.com/video/gemini-robotics2-transfer-cost-vla/https://frontieraianalysis.com/assets/hover-gifs/gemini-robotics2-transfer-cost-vla.jpgGemini Robotics 2: Can a new robot reuse another robot's skills?Google DeepMind’s official whole-body robotics demonstration. A company demonstration, not independent validation.https://storage.googleapis.com/gdm-deepmind-com-prod-public/media/MVCfTE-4h3AwDVPZ/GR2_Wholebody_posterframe_v01.webm2026-07-30T16:00:00+00:0010https://frontieraianalysis.com/video/physical-intelligence-pi07-steering-models/https://frontieraianalysis.com/assets/hover-gifs/physical-intelligence-pi07-steering-models.jpgPhysical Intelligence's π0.7: Teaching new tasks through better instructionsPhysical Intelligence’s π0.7 demonstration of cutting a zucchini. The publisher’s clip illustrates the task; it is not a general success-rate measurement.https://website.pi-asset.com/pi07/cut_zucchini_compressed.mp42026-04-1674https://frontieraianalysis.com/video/mecka-human-demonstrations-model-training/https://frontieraianalysis.com/assets/hover-gifs/mecka-human-demonstrations-model-training.jpgIntroducing MeckaMecka’s company video illustrates its approach to human demonstration data; it is not a validation of transfer to every robot.https://www.youtube-nocookie.com/embed/7ECZh8qP_t82026-06-0191https://frontieraianalysis.com/video/snorkel-data-environments-model-curriculum/https://frontieraianalysis.com/assets/hover-gifs/snorkel-data-environments-model-curriculum.jpgCoastRunners 7The video accompanying Snorkel’s discussion of reinforcement-learning environments, rewards and validation.https://www.youtube-nocookie.com/embed/tlOIHko8ySg2016-12-2157https://frontieraianalysis.com/video/datology-curation-studio-training-budget/https://frontieraianalysis.com/assets/hover-gifs/datology-curation-studio-training-budget.jpgIntroducing Datology Curation StudioOfficial Curation Studio video embedded in Datology’s product announcement.https://www.youtube-nocookie.com/embed/Zd527P1qRUI2026-10-09154https://frontieraianalysis.com/video/flashattention-four-training-speed-experiment-budget/https://frontieraianalysis.com/assets/hover-gifs/flashattention-four-training-speed-experiment-budget.jpgLecture 80: How FlashAttention 4 WorksCharles Frye explains the FlashAttention 4 forward-pass implementation on Blackwell GPUs. This developer lecture predates the 2026 research paper and supplies implementation context, rather than a validation of that paper’s results.https://www.youtube-nocookie.com/embed/VPslgC9piIw2025-10-014509https://frontieraianalysis.com/video/dflash-two-speculative-decoding-verification-economics/https://frontieraianalysis.com/assets/hover-gifs/dflash-two-speculative-decoding-verification-economics.jpgDFlash 2: Keep Drafting ParallelInco AI’s official DFlash 2 video, as embedded in its release announcement.https://www.youtube-nocookie.com/embed/s2-GRNDn_M82026-08-1827https://frontieraianalysis.com/video/prism-bonsai-two-compression-deployment-boundary/https://frontieraianalysis.com/assets/hover-gifs/prism-bonsai-two-compression-deployment-boundary.jpgPrismML: A smaller model could run on hardware you already ownPrismML’s 59-second demonstration of a skateboard game created with Bonsai 2 and the Hermes agent. This is one recorded demo, not a general coding benchmark.https://raw.githubusercontent.com/PrismML-Eng/Bonsai-demo/main/demos/skateboard/skate-runner/demo_clip_seed42_59s.mp459https://frontieraianalysis.com/video/mercor-human-baselines-autonomy-accounting/https://frontieraianalysis.com/assets/hover-gifs/mercor-human-baselines-autonomy-accounting.jpgMercor: Winning an accounting test does not prove lower workplace costsMercor’s introduction to its accounting productivity index, providing context for the human-baseline comparison.https://player.vimeo.com/video/1214103504?h=3e64e114122026-07-31T09:30:00.000-07:00180