OPEN LINK ↗
// 60d agoRESEARCH PAPER
Price Is Not Right slashes energy, tops VLAs
Tufts and AIT researchers benchmarked a neuro-symbolic robotics stack, combining PDDL-based symbolic planning with learned low-level control, against fine-tuned OpenPi π0 VLA models on Towers of Hanoi in simulation, with the paper accepted for ICRA 2026. The hybrid method hit 95% success on the 3-block task, generalized to an unseen 4-block variant, and used roughly two orders of magnitude less training energy.
// ANALYSIS
This is a strong reminder that structured robotics doesn't always reward bigger end-to-end models. When the task has explicit rules, decomposition can buy both reliability and a much saner power budget.
- –The neuro-symbolic model trained in 34 minutes and about 0.85 MJ, while the VLA fine-tunes took roughly 1d16h and 64.9-68.5 MJ.
- –Inference was cheaper too: the NSM used 19.4 W total and no GPU in the benchmark, versus about 114-115.2 W for the VLA setups.
- –The 4-block Hanoi result is the real signal here: 78% success on an unseen variant versus 0% for both VLA baselines points to rule transfer, not demo memorization.
- –The planner-guided VLA still failed on the 3-block task, which suggests symbolic hints bolted onto a VLA are not the same thing as a true symbolic planner.
- –The open code and models on GitHub and Hugging Face make this a useful follow-up baseline for anyone working on VLA or neuro-symbolic robotics.
// TAGS
the-price-is-not-rightroboticsreasoningresearchopen-source
DISCOVERED
60d ago
2026-03-30
PUBLISHED
60d ago
2026-03-30
RELEVANCE
8/ 10
AUTHOR
callmeteji
