Discovering State and Action Abstractions for Generalized Task and Motion Planning
Authors: Aidan Curtis, Tom Silver, Joshua B. Tenenbaum, Tomás Lozano-Pérez, Leslie Kaelbling5377-5384
AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Lastly, we evaluate our algorithm s ability to scale to large unseen problems and compare planning speeds with a standard TAMP solver. Additionally, we show that these simple generalized plans learned from only a handful of examples can be used to improve the search efficiency of TAMP solvers. As seen in Figure 3, our experiments show that execution speed from the GENTAMP policy greatly outperforms a state-of-the-art TAMP algorithm. The difference in planning speed increases exponentially with object count. The bottom four panels show planning speed comparisons between our approach, standard TAMP, and partial skeleton. Each trial contains six seeds used to randomize object type, placement, size, and color. The plot error bars show a 95% confidence interval. |
| Researcher Affiliation | Academia | Aidan Curtis, Tom Silver, Joshua B. Tenenbaum, Tom as Lozano-P erez, Leslie Kaelbling MIT Computer Science and Artificial Intelligence Laboratory {curtisa, tslvr, jbt, tlp, lpk}@mit.edu |
| Pseudocode | Yes | Algorithm 1: GENTAMP Training Input: P = {P1, P2, . . . , PN 1} Q Output: πφ D = {}, pool {} for Pi P do D D TAMP(Pi) end for c {1 . . . max complexity} do pool PRUNE(pool GENERATE(pool)) end T ENCODESAT(D, pool) F, A SATSOLVE(T ) I, G F(P) πφ FONDPLAN(F, A, I, G) return πφ, F |
| Open Source Code | No | The paper discusses using existing tools like PDDLSTREAM and Open WBO but does not state that the authors' implementation code for GENTAMP is open source or provide a link to it. |
| Open Datasets | No | Our approach uses a TAMP solver to find example plans for each training problem instance and uses the transitions from those plans to create approximately valid abstractions. The paper describes generating training examples but does not provide access information (link, DOI, or citation) for a publicly available dataset. |
| Dataset Splits | No | The paper refers to 'training problem instance' and 'testing problem instances' but does not specify exact split percentages (e.g., 80/20) or a cross-validation methodology for data partitioning. |
| Hardware Specification | No | The paper does not explicitly describe the specific hardware (e.g., CPU, GPU models, memory) used to run the experiments. |
| Software Dependencies | No | The paper mentions external tools like PDDLSTREAM, Open WBO (Martins, Manquinho, and Lynce 2014), and PRP (Muise, Mc Ilraith, and Beck 2012) but does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | Yes | We prune concepts and quantified concepts in each step by uniqueness and equivalence. The specifics of this generative grammar are in Table 1. Each feature in this grammar is accompanied with a feature complexity score that is defined as the maximum of the number of generative grammar rules applied to create the feature and the number of arguments of the feature. This grammar is implemented in a bottom-up fashion by combining primitives, concepts, and quantified concepts to create a full feature and discarding features when they pass a fixed complexity limit (5 in our experiments). Each trial contains six seeds used to randomize object type, placement, size, and color. |