A robot policy reached a reported 91.7% task-success rate after researchers trained it to connect each move with the reason for taking it, up from 9.5% for the original policy in the same hardware benchmark. The result comes from a small lab evaluation, but it points to a way to improve existing robot models without collecting new demonstrations.
Image: Figure 1 from the ARC paper, by its authors, reproduced unmodified under CC BY 4.0.
Robot policies commonly learn by copying demonstrations: they map camera images and an instruction to a sequence of movements. The demonstrations show what the robot did, but rarely explain why a particular move was right. ARC, a recipe from researchers at several universities and NVIDIA, adds a short natural-language trace alongside each action chunk. The trace identifies the relevant scene state, why it matters, the intended effect, the next action, what to avoid and whether the task is complete.
The team generated these traces from the existing DROID robot-demonstration dataset rather than recording new demonstrations. The paper reports about 75,000 relabeled episodes and 1.2 million frames aligned with actions. It then fine-tuned two existing systems: pi0.5, a vision-language-action model, and Cosmos3-Nano-Policy, a model that predicts actions alongside future observations. At run time, an external vision-language model writes the trace; the trained robot policy still produces the continuous movements.
In simulation, ARC raised pi0.5's success on RoboLab-120's vague-instruction tasks from 15.2% to 45.1%. The physical test used the pi0.5 policy on the paper's DROID setup, a Franka Panda arm with a two-finger gripper. ARC reached 91.7% versus 9.5% for the base policy, an 82.2-percentage-point difference. The authors used three episodes per task and did not fine-tune the model on hardware-specific data.
The result is promising but narrow: it is an author-run test of one robot embodiment, not evidence that the policy will generalize across hardware or unfamiliar settings. ARC also depends on the vision-language model reading the scene correctly. The paper warns that occlusion can lead it to mistake an unfinished grasp or placement for a completed step, and that an old reasoning trace can remain in use while a new one is generated. That adds computation, memory use and latency as well.
The paper was posted to arXiv on October 8. The authors say they plan to release the models, code and ARC-Trace dataset after the review period.