In simulated kitchen-table organization tasks, a planner that encoded inter-object affordances - constraints on how objects can support one another during grasping, placement, sliding and lifting - reported much higher success rates than the original U-TAMP and a vision-language model (VLM) task planner. In pick-and-place-only configurations, UTAMP-2-GT, the version using hand-crafted ground-truth affordances, reached 100% success. UTAMP-2-VLM, which used VLM-detected affordances, reached 90%; the other approaches were below 50%, and the VLM planner was reported at 39%.
The benchmark produced 127 valid stack combinations, with the number of manipulable objects varying from 2 to 5. Success rate was calculated as the share of initial configurations that produced successful simulated execution, while planning computation time was reported as a mean and standard deviation across initial configurations. The document is a preprint identified as arXiv:2608.25641v1 and dated 26 August 2026.
Rules that capture object relationships
UTAMP-2 is a PDDL-based extension of U-TAMP. In practical terms, the approach expresses inter-object affordances as explicit constraints for grasping, placement, sliding and lifting. The PDDL-based approaches used the off-the-shelf Fast Downward planner.
The comparison included original U-TAMP, known as UTAMP-1, UTAMP-2-GT with hand-crafted ground-truth affordances, UTAMP-2-VLM with VLM-detected affordances, and a VLM task planner using common-sense reasoning. The motion-planning step used RRT-Connect in CoppeliaSim, while the VLM used in the evaluation was OpenAI's o4-mini with high reasoning effort.
The simulated scene used CoppeliaSim with a Franka Emika Panda robot, a Panda gripper and five kitchen objects on a table.
The gap held when sliding was required
One of the study's clearest warnings came from the VLM-assisted version. Its mean precision, recall and F1 scores for affordance detection were all 0.98, yet its task success was 90%. The authors linked the failures to a few incorrectly assigned supporting affordances. High aggregate detection scores therefore did not correspond to perfect end-to-end task success in this benchmark.
Planning time created an equally stark divide. The UTAMP-2 variants planned about two orders of magnitude faster than the other methods. VLM planning was, on average, two orders of magnitude slower than UTAMP-2 and had the greatest variance. As complexity increased, UTAMP-2 retained higher success, while VLM and UTAMP-1 seldom succeeded when five objects were involved.
The difference was also visible in configurations that required sliding. UTAMP-2-GT had a success rate of 1.0, UTAMP-2-VLM had a rate of 0.81, and VLM had a rate of 0.41. Their mean planning times were 0.17 seconds, 0.20 seconds and 66.9 seconds, respectively. The corresponding standard deviations were 0.05, 0.08 and 25.2 seconds.
A result bounded by its simulation
The authors interpret the comparison as showing that robust symbolic abstractions of inter-object affordances are associated with quickly generated, feasible and expressive plans without sub-symbolic reasoning. The paper also acknowledges that its affordance and motion-planning assumptions are simplified and may not hold for all objects and scenarios.
The evidence is bounded by one simulated kitchen-table scenario with five specified objects; no evaluation of other environments, object sets or physical robots is reported.
The report gives no confidence intervals, hypothesis tests or p-values. It also does not provide counts for each task subset and planner, leaving the reported comparisons as benchmark summaries rather than formal inferential results.
Paper data and sources
Original title: Leveraging Inter-object Affordances for Efficient Planning in Contact-rich Tasks
Authors: Pouya P. Niaz, Justus Piater, Alejandro Agostini
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text