Robotics models often struggle outside controlled environments. Ours is built to work in real ones. Today we're launching MolmoAct 2, which can assist with a host of chores & lab tasks, plus the MolmoAct 2-Bimanual YAM dataset—the largest open robotics dataset of its kind. 🧵

lesgoo i’m still so amazed with molmoact2 last week i recorded a few hundred episodes and tried to fine tune all the vlas on lerobot molmoact2 was simply the undisputed best, no other vlas can come close the only thing that came close was multitaskdit, but it’s not a vla, and is small enough to be trained from scratch and be sample efficient i’ll do a write up on this soon, also going to leverage the molmoact2 dataset to reproduce their benchmark some folks told me that molmo’s moat was only data on the so101, had pi0.5 been trained with molmoact’s so101 dataset, it would have generalized too i simply don’t think that’s true (from the paper’s claims), but we’ll see

Tried to get MolmoAct2 to zero-shot pick up a cube with an SO101. The shoulder lift servo just slammed straight into the table. Finally got it working with @pham_blnh's help. The checkpoint is trained in old SO-100-era degrees (pre-PR #777). Newer SO-101 calibration flips offsets the shoulder/elbow zero and normalizes units. If you try using the new calibration on the policy zero shot, arm slams down instead of lifting. Fixes are in this @LeRobotHF fork: github.com/makermods-robotic…

Cosmos 3 wins another leaderboard. The power of an open frontier model. 🎉 Ranked #1 on @allen_ai's MolmoSpaces robot policies leaderboard, a policy model post-trained on Cosmos 3 Nano shows the strongest performance in environments that require models to interpret, reason, and act in real time. Not just understanding the world, but interacting with it. 📊 molmospaces.allen.ai/leaderb… Drop what you're building in the comments 👇

Since MolmoAct 2 works zero shot; sims are good place to start experimenting with models, RL and other such unconventional tricks. Here it took me couple of hours to get Genesis World working with MolmoAct2 checkpoint for bimanual YAM setup.

MolmoAct (zeroshot) meets auto evals. Thanks @DJiafei for the suggestion. What should I build next? A) Train a policy to dump the cup (no custom hardware) B) Collect intervention data to push policy performance C) Train a simpler task like pressing buttons D) Something else

my learnings from reading the paper if anyone is interested aryanmadhavverma.com/posts/r…

There are few truly open models in the world, including both weights and data. However, these models are crucial for research and development of new systems — they help us learn which data is important and help develop new capabilities for deploying robots in the real world. MolmoAct2 provides a foundation for open research into robotics. It is associated with its own open dataset, an open-data action tokenizer, and a reasoning variant which predicts depth tokens. And people have actually been using it across the community, running experiments in their own labs or homes. @hq_fang and @DJiafei tell us more. Watch Episode 87 of RoboPapers, with @micoolcho and @chris_j_paxton, now!

There are few truly open models in the world, including both weights and data. However, these models are crucial for research and development of new systems — they help us learn which data is important and help develop new capabilities for deploying robots in the real world. MolmoAct2 provides a foundation for open research into robotics. It is associated with its own open dataset, an open-data action tokenizer, and a reasoning variant which predicts depth tokens. And people have actually been using it across the community, running experiments in their own labs or homes. @hq_fang and @DJiafei tell us more. Watch Episode 87 of RoboPapers, with @micoolcho and @chris_j_paxton, now!

We're releasing MolmoMotion, a 3D motion forecasting model. Given one or a few video frames, 3D points on an object, & an instruction like "Put the white bowl on the table," MolmoMotion predicts where those points will go over the next few seconds in a shared 3D world frame. 🧵

Announcing @xdofai: We’ve raised $70 million to build the core robotic infrastructure ecosystem for robot foundation models. My cofounders Fred (@YideShentu), Nemo (@itsnemojin) and I have been pursuing the dream of general purpose robots for our entire lives. After work at Covariant, Meta and Tesla, it became clear to us that general purpose robots are coming, and we are building XDOF to help make them a reality. For the last two years, we’ve been working behind the scenes to support major labs and companies deploying robots. In us, they have a partner with full-stack expertise, from hardware to operations to policy training. As our first public contribution to the space, we are open-sourcing ABC-130K, the largest open source teleoperation dataset, in collaboration with our partners from UC Berkeley, Carnegie Mellon, MIT and Amazon FAR. Thank you to our customers, partners, collaborators and investors for your trust and conviction in us. Together, we can accelerate the future of robotics!

Yams running Momo entirely from the cloud. Robots in nature can't afford to carry their compute on their back. Cloud inference can flip it. Moving insanely fast in the lab, been quiet going full peak mode. More to come soon.

Autoresearch just left the sandbox and entered the embodied world. We are excited to introduce 𝐄𝐍𝐏𝐈𝐑𝐄: a system that drops frontier coding agents onto a fleet of real robots and hands them the entire loop: reset the environment → search the literature → implement ideas and build the infra → train and deploy → self-verify → analyze the logs and rewrite the code → repeat, until the policy is reliable in the real world. No human in the loop. Guided only by the robot's self-proposed, heuristic-based success signal, the agents hill-climb to 99% on dexterous real-world tasks: organizing pins into a box, seating GPUs, tying zip-ties. We envision the bottleneck in robotics shifting — from building smarter algorithms to building the closed physical feedback loops an agent can finally turn on its own. 🔗 research.nvidia.com/labs/gea… From @NVIDIA @CMU_Robotics @Berkeley_AI 🧵

What happens when actions fail—and VLAs don’t know how to recover, especially when they’re frozen? We introduce corrective decoding: bootstrapping a pre-trained VLA with a companion VLM, FailSafe trained only from simulation, that detects impending failures and issues executable 7-DoF fixes. More reliable manipulation, fewer dead-ends. 🔧🤖 jimntu.github.io/FailSafe/