Axis Robotics
516 posts
@axisrobotics
Scale Physical AI for the real world. Robot intelligence is not built by a few; it's built by all.
Want to know how Access Codes fuel real ecosystem growth beyond @BitRobotNetwork? 💡Join @CryptitaPlays & @AxisRobotics for an exclusive AMA on gateway rewards & AirDrop allocations! 🗓️ July 22, 2026 | 9 PM PHT 👉 Listen here: #AxisAccessCode" target="_blank" rel="nofollow noopener">t.me/cryptitaplays#…
#Web3
Axis Weekly This week, we turned the direction we set for DAgger and TaskGen in June into more concrete execution work. We moved the correction-data pipeline toward training, prepared the post-task deployment workflow, improved TaskGen and rendering infrastructure, brought OpenArm online, and made object-selected gripper movement smoother. Key updates: - DAgger: DAgger (Dataset Aggregation) is a training approach that improves a policy using corrections collected during policy rollouts. This week, we defined an initial scoring method and began preparing a controlled small-model setup for evaluating correction data. - Post-task deployment: Training, packaging, uploading, testing, and release preparation are being connected into a repeatable workflow. - TaskGen: We identified key bottlenecks and are working to separate, automate, and parallelize layout construction and asset generation. - Infrastructure and benchmarks: Parallel rendering roughly doubled throughput in the current setup, while Libero Pro migration is being prepared. - OpenArm and control: OpenArm was brought online, and object-selected gripper movement was made smoother so the resulting trajectories are easier for policies to learn from. Details below ⬇️
On the DAgger side, the pipeline moved from infrastructure work toward training execution. DAgger, short for Dataset Aggregation, improves a policy using corrections collected while the policy is running. We are now exploring how to score these collected trajectories based on three factors: trajectory quality, policy-control time, and total control time. One initial formulation is: dagger_score = length.score × policy_control_time / total_control_time This adjusts the trajectory-quality score based on how much of the total rollout remained under policy control. The next step is to work through the remaining training tasks and reproduce a state-based diffusion policy baseline. This will provide a controlled small-model setup for evaluating correction data. Read our full thesis on DAgger and Axis V2 here: axisrobotics.ai/resources/prod…
lots of conversations about base over the last week. wanted to share my candid take after a week of listening and a lot of reflection over the last 6 months. first off - in case it’s not obvious, the first quarter of 2026 was a punch in the face. I spent 2024 and 2025 making a two pronged bet to bring base to the world: (1) builders would unlock the next wave of crypto adoption; (2) adoption would be driven by new onchain-native social experiences - creators, content, messaging. imo we made the right bet on builders, but obviously the wrong bet on social. builders did drive the next wave of crypto adoption - prediction markets, perpetuals, stablecoins - but social was not at the center of it. in fact, the entire social side of the market that many of us had been building towards - farcaster, zora, miniapps, and yes, creator coins - disintegrated completely. I was wrong - whether it was timing wrong (is $ansem a creator coin?) or fully wrong, only time will tell, but regardless, i was definitively wrong. the collateral damage was pretty bad! and this year has been an exercise in eating shit. we realized how our focus on social had meant that base had fallen behind in key areas that were now increasingly critical - we had perps (shoutout avantis!) and prediction markets (shoutout limitless!), but both were well behind scaled competitors. and we had a lot of room to improve in unlocking base as a platform for tokenization and payments that really worked for enterprises. people lost confidence, and CT spectators reminded me weekly of all of my mistakes as often as they could. it felt bad man, still feels bad. but if there’s one thing i’ve learned from the last decade of building in this space, it’s that when things feel the worst, the best thing to do is just put your head down and build. so that’s what i’m doing. I refocused my time and attention back to the chain away from the app, started writing code again, shipped a bunch of stuff (azul, beryl, b20, privacy, ledgers) and questioned a bunch of my assumptions: does crypto need social to grow? does base need an app? can base be bigger than coinbase? I thought for a long time that social was the only thing that could drive the sort of viral growth to get crypto to a billion people. unsurprisingly, I now believe that’s wrong. It’s clear that better money is more than enough - we are seeing this live with stablecoins, predictions, perpetuals, tokenization and i only expect it to accelerate. I am now focused on bringing a billion people onchain just by making global finance actually work. on the app, my focus is on building base into the blockchain for global finance. to that end, i’ve handed the base app back to the coinbase mothership, where my now good friend @cobie will be taking it from here to make it the best damn app for onchain you’ve ever seen, including expanding beyond the base ecosystem in ways that tbh i won’t love as the leader of base. it’s incredibly hard to grow a decentralized network inside of a big public corporation. and i feel like much of the discourse on CT over the last week is downstream of this. the following things can be true: (1) base (and i) love memes and (2) brian probably won’t ever bullpost memes on the tl (this activity is illegal once you’re over 40 years of age). it’s weird and we’re working through it as we continue to decentralize base, which has been our commitment from the beginning. we’re going to build base into the blockchain for global finance and do everything we can to be the place that the world’s money settles over the next century. we will surely have formidable competitors (welcome robinhood and stripe!) and people may abandon our cause, but we welcome the competition and believe it’s our duty to win the respect and commitment of those who rally to our banner. in 2026, this concretely means three things: winning trading, payments, and agents. [continued in the reply]
River(@RiverdotInc )の素晴らしいチーム、そして東京でご参加いただいた皆さま、本当にありがとうございました！🇯🇵 皆さまのサポートのおかげで、今回のイベントは最高のコミュニティイベントの一つとなりました。 Physical AIコミュニティを共に盛り上げてくださったすべての皆さまに感謝します。 また次のイベントでお会いしましょう！🚀
