Sequential Planning via Anchored Robotic Keypoints
Sequential Planning via Anchored Robotic Keypoints
This paper introduces Sequential Planning via Anchored Robotic Keypoints (SPARK), a framework that decouples symbolic task planning from dynamic scene observations. By anchoring actions to current 3D data at runtime and verifying plans locally, the system recovers from object displacements without re-planning. Tested on three different robots with no prior practice or simulator access during rollouts, SPARK achieves a mean success rate of 66.8% across various position and task settings using an open-weight 27B planner with 4-bit weights.
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
position and task settings
6
View original evidence
retry. We use one open-weight 27B planner with 4-bit weights. Under this protocol \spark achieves a mean success of $66.8\%$ across the 6 position and task settings and $66.8\%$ on the held-out language and object settings. We also achieve a $68\%$ success rate across 11 physical task-embodiment settings on 3 different robots. Our work demonstrates that for an important class of failures the robot does not need to a
Open source S5language and object settings
1
View original evidence
retry. We use one open-weight 27B planner with 4-bit weights. Under this protocol \spark achieves a mean success of $66.8\%$ across the 6 position and task settings and $66.8\%$ on the held-out language and object settings. We also achieve a $68\%$ success rate across 11 physical task-embodiment settings on 3 different robots. Our work demonstrates that for an important class of failures the robot does not need to a
Open source S5physical task-embodiment settings
11
View original evidence
retry. We use one open-weight 27B planner with 4-bit weights. Under this protocol \spark achieves a mean success of $66.8\%$ across the 6 position and task settings and $66.8\%$ on the held-out language and object settings. We also achieve a $68\%$ success rate across 11 physical task-embodiment settings on 3 different robots. Our work demonstrates that for an important class of failures the robot does not need to a
Open source S5
- code: Reported pending Open source S6
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
cuting. When displacing objects by $10$\,cm we were still able to complete $66.7\%$ of trials with scene-diff verification and layered recovery, and we did not have to change the plan. We test our full system on \libpro with these conditions. There is no practice on the tasks being tested and no access to simulator states during rollouts, and as in \rats and \capx a success check after each attempt decides whether to
Open source S4
retry. We use one open-weight 27B planner with 4-bit weights. Under this protocol \spark achieves a mean success of $66.8\%$ across the 6 position and task settings and $66.8\%$ on the held-out language and object settings. We also achieve a $68\%$ success rate across 11 physical task-embodiment settings on 3 different robots. Our work demonstrates that for an important class of failures the robot does not need to a
Open source S5
sk for a new plan for its task. It can be enough to update the system's view of the world. Code and videos can be found at https://cwru-aism.github.io/spark-page/
Open source S6
Source：arXiv Robotics — research abstracts · arxiv.org