Robotics work centers on real-world long-horizon evaluation

LongBench gives this period a concrete anchor: long-horizon robot evaluation is getting more specific about why policies fail. The benchmark covers 10 real-world tasks and more than 1,000 episodes. It separates fully observable execution problems from context-dependent ambiguity, then scores progress stage by stage instead of using a single pass/fail number. That matters because current policies break for different reasons. On context-independent tasks, pi_0 leads with an average stage-wise score of 86.3, while Diffusion Policy reaches 51.2 and OpenVLA-OFT 32.7. Dynamic grasping is the clearest stress case. pi_0 gets 73.3, while the other listed systems stay between 0.0 and 13.3.

The second signal is broader than one paper. The robotics stack in view here is tied to data scale, real deployment, and multimodal action prediction. The Technology Review history piece is not a benchmark paper, but it gives useful context for why long-horizon evaluation matters now. It cites Google RT-1 training over 17 months and 700 tasks, with 97% success on seen tasks and 76% on unseen instructions. It also notes growing commercial pressure: humanoid robot investment reached $6.1 billion in 2025. Read together with LongBench, the near-term emphasis is clear. Better robot models need better real-world tests, because larger policies and larger data collection loops are already in deployment discussions.