This Week in
Robotics
A weekly technical briefing on policy learning, manipulation and embodiment. Mechanism over hype.
Issues
Issue No. 8
Touch stopped being an input this week. Two results, pointing from opposite ends at the same conclusion: the value of contact data is in the training signal, not in the machinery you build to consume it. Spark-0 (preprint, Xspark AI + THU + MMLab@HKU + HKUST-GZ + PKU) uses 1,000
Issue No. 7
Last week three groups converged on how a predicted future acts on a policy. This week four groups independently measured where that whole program stops, and every ceiling they found is about data or representation rather than architecture. RoboJEPA (preprint, FAIR at Meta + Mila
Issue No. 6
Three independent groups converged on the same answer to a question the field has been arguing about for two quarters: the predicted future acts on a world-action policy through its presence and structure, not through its content. Simple-WAM (preprint, Tsinghua Leap Lab) shows th
Issue No. 5
The infrastructure moved more than the models. Alphabet’s Intrinsic open-sourced its production arm stack under Apache-2.0 and, a day later, Qualcomm agreed to buy PickNik, the company that maintains MoveIt 2 — the open manipulation stack changed hands and gained a major do
Issue No. 4
The week’s most useful paper tore down an old result instead of adding a new one. A forensic re-run of ACT’s CVAE ablation finds the latent carries no information at default settings and the famous 35%→2% drop does not reproduce. Around it, force made its strongest ca
Issue No. 3
The tooling layer moved for the first time since spring — LeRobot v0.6.1, MuJoCo 3.13.0 with a new discrete integrator, Genesis 1.4.1 — and OpenWAM turned last week’s uncarded checkpoint dump into a fully released, controlled study of world-action-model pretraining, while t
Issue No. 2
NVIDIA agreed to buy Hugging Face for $12.9B, putting LeRobot’s maintainer inside the company that ships Isaac GR00T — and in the same week a 0.54M-parameter policy with no language encoder and no pretrained vision hit 95.1% on LIBERO, which says more about the benchmark th
Issue No. 1
A 0.68M-parameter predictive-coding policy from Sony CSL beat parameter-matched Transformer and LSTM baselines by 4–7× on LIBERO, which is a direct attack on the premise that language-conditioned control needs VLA-scale parameters — while capital moved the opposite way, with Soft