Can a Frontier VLM Serve as a Robot Policy for Memory-Dependent Manipulation?
Can a Frontier VLM Serve as a Robot Policy for Memory-Dependent Manipulation?
This research shows that a frozen frontier vision-language model, GPT-6-Astra, can act as a robot policy for memory-dependent manipulation tasks on the RoboMME benchmark. The model achieves a 92.1% success rate on 800 test episodes, outperforming human performance and existing benchmarks. The results suggest that a single general-purpose model may take over significant aspects of embodied systems, including long-horizon planning, memory, and end-effector control.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
test episodes
800
View original evidence
arXiv:2610.17350v1 Announce Type: new Abstract: We use a frozen frontier vision-language model, GPT-6-Astra, directly as the robot policy on RoboMME, a benchmark of 16 memory-dependent manipulation tasks spanning temporal, spatial, object, and procedural memory. The model keeps the whole episode in context and outputs every end-effector waypoint and gripper command, with no RoboMME-specific training. On the 800 test
Open source S2tasks matched or exceeded
13
View original evidence
episodes it succeeds in 92.1%, above the previously reported human score of 90.5%, which was obtained with an oracle skill interface, and far ahead of the best leaderboard entry (60.8%) and of a hierarchical system that pairs the same model with a RoboMME-trained policy (79.1%). It matches or exceeds the human score on 13 of the 16 tasks. Its 63 failures concentrate in two tasks and are mostly premature completions a
Open source S3failures
63
View original evidence
episodes it succeeds in 92.1%, above the previously reported human score of 90.5%, which was obtained with an oracle skill interface, and far ahead of the best leaderboard entry (60.8%) and of a hierarchical system that pairs the same model with a RoboMME-trained policy (79.1%). It matches or exceeds the human score on 13 of the 16 tasks. Its 63 failures concentrate in two tasks and are mostly premature completions a
Open source S3
- code: not_reported
- paper: not_reported
- dataset: not_reported
- weights: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
arXiv:2610.17350v1 Announce Type: new Abstract: We use a frozen frontier vision-language model, GPT-6-Astra, directly as the robot policy on RoboMME, a benchmark of 16 memory-dependent manipulation tasks spanning temporal, spatial, object, and procedural memory. The model keeps the whole episode in context and outputs every end-effector waypoint and gripper command, with no RoboMME-specific training. On the 800 test
Open source S2
episodes it succeeds in 92.1%, above the previously reported human score of 90.5%, which was obtained with an oracle skill interface, and far ahead of the best leaderboard entry (60.8%) and of a hierarchical system that pairs the same model with a RoboMME-trained policy (79.1%). It matches or exceeds the human score on 13 of the 16 tasks. Its 63 failures concentrate in two tasks and are mostly premature completions a
Open source S3
Source：arXiv Robotics — research abstracts · arxiv.org