Axis Weekly This week, we continued strengthening our closed-loop robotics data pipeline, from TaskGen and simulation infrastructure to failure recovery and asset-level augmentation. Key updates: - Task generation: We completed asset scan and merged it into TaskGen, helping generated tasks reason over available assets, scene layouts, long-horizon workflows, and multi-embodiment settings. - Simulation infra: We improved MuJoCo verify, replay, and scene-variant workflows, with fixes around repeated downloads, caching, compatibility, and long-horizon multi-asset task stability. - Robot controls: We cleaned up gripper behavior, IK, teleoperation, and the control panel based on feedback from longer-horizon and multi-asset tasks. Failure recovery: We continued building a pipeline to turn failed and near-failed grasping states into reusable data for recovery learning. - Asset augmentation: With academic collaborators, we advanced a shape augmentation direction that can expand one seed asset into many physically plausible object variants. A closer look at this week’s progress 🧵

.@Figure_robot’s 100-hour sorting marathon just showed a human worker narrowly beating an autonomous robot, even as his arm nearly gave out. Figure is demonstrating the massive potential of intelligent robots in physical production, while Axis is building the underlying infrastructure to support and scale this robotic intelligence. Robots like these may not beat humans in every direct contest yet. But on Axis, your data can help train them, advance them, and bring them closer to that future. In that sense, the student may one day surpass the master.

Yesterday our founder @chris_anm01 joined the community for an AMA in our Discord channel. We’ve shared a lot on X about the engineering behind our data engine, but this session went much deeper. Chris broke down our actual competitive moat, our commercial roadmap, and the long-term vision for Axis—critical details we haven't fully unpacked here yet. Here are the key takeaways you need to know. 🧵

Axis Robotics Philippines 🇵🇭 – First Ambassador Line-up A strong community is never built by one person alone, but by individuals who believe in the future of Physical AI and are committed to creating long-term impact together. Today, @axisrobotics proudly introduces its first 3 Official Ambassadors for the Philippines community: @MonetaGratia15 — The Live Streamer @Belzky2 — The Referral Queen @Chunkyweb3 — The Content Creator These ambassadors represent the beginning of Axis Robotics’ journey toward strengthening and expanding the Physical AI movement within the Philippines 🇵🇭 More than simply sharing information, they will play an important role in educating, connecting, and empowering the community while helping grow the Physical AI ecosystem locally. And this is only the beginning. Keep contributing. Keep building. Keep growing with the community because the next Axis Robotics Ambassador could be you. #AxisRoboticsPH#

We're building a Task Generation System for robot learning. Four design principles from day one: Structured — Tasks are defined through clear data hierarchies and parameterization, not loose script stacking. Productized — This isn't an internal tool. It's a configurable, deliverable product that non-engineers can understand and operate. Extensible — New scenarios, new capabilities, new asset classes — all additive, never requiring a rebuild. API-first — The end goal is open access. Interface design, parameter specs, and output formats are built for external consumption from the start. This is the engine behind data diversity

Physical AI should be built by everyone. Proud to see 500+ students in the Philippines experience Axis firsthand and contribute to the future of robotics data generation. Big thanks to @basepilipinas for making this possible and bringing such amazing energy to the community 💙 More student roadshows and campus activations across APAC with @baseapac coming soon!

Axis Weekly Last week, we made progress across the full robotics data loop, including task generation, simulation infrastructure, model training, and failure recovery. Key updates: - Task generation: We improved TaskGen with better automatic checker generation, stronger multi-embodiment support, and more efficient domain randomization to scale task diversity with less manual design effort. - Simulation infra: We continued improving MuJoCo verify/replay and scene-variant workflows, including fixes across data collection, multi-asset scenes, repeated loading/downloads, initial states, teleoperation, IK, and gripper control. - Model training: We confirmed that the new randomized tasks are learnable with sufficient data. In our current experiment, 500 demos successfully produced an executable policy, while 100 demos were not enough. - Failure recovery: We began building a recover-from-failure pipeline to collect and categorize gripper failure and near-failure states during grasping, which will later support more robust recovery policy learning. A closer look at this week’s progress🧵

We are now open-sourcing the AxisDataCleaning pipeline. Github repo: Browser teleoperation is one of the most scalable paths for robot data generation. Raw human input, however, is not yet model-ready: ▪️ Idle pauses ▪️ Micro-jitters ▪️ Low & Variable frame rates Raw web data alone is not enough for reliable policy training. Here is how our backend turns noisy human demonstrations into usable trajectories for downstream policy training. 🧵👇