VLA action choice stops trusting the top score alone: FRS-Grounding adds neighborhood consistency for 86.4% simulated success
Synopsis
The highest-scoring action from a vision-language-action (VLA) model is not always the most reliable one to execute. FRS-Grounding keeps the finite candidate set after VLA inference, fuses normalized candidate scores with distance to a safe-action set into a fuzzy action set, and builds a pairwise relation from pose differences and collision-free states before fuzzy rough lower and upper approximations separate strongly supported candidates from boundary cases. Across 1500 randomized simulations the full pipeline reached 86.4% success and a 2.1% safety-violation rate, versus 74.8% and 8.2% for the pointwise hard-constraint baseline VLA-HC; under positional noise with σ = 10 mm it retained about 78% success against about 28% for VLA-HC.
Interpretation
Action selection moves from scoring each candidate alone to evaluating each action together with whether its nearby actions agree. Earlier pointwise practice ranked candidates independently by VLA-conditioned score, or applied collision and workspace hard constraints to individual candidates, so the final decision rested only on that candidate's own state or cost. On 1500 matched randomized tabletop scene-instruction instances, the complete FRS-Grounding pipeline reached 86.4% success and a 2.1% safety-violation rate, against 74.8% and 8.2% for VLA-HC operating on the same candidate representation.
Each candidate first receives a semantic-physical membership, and candidate pairs then receive graded relation values from pose proximity and collision-free status. The semantic term is the normalized VLA-conditioned score; the physical term is an exponentially decayed support value from the minimum normalized pose distance to the reference safe-action set, and the two are combined linearly with weight α = 0.70. Starting from the OpenVLA-7B nominal action, a six-dimensional Sobol sequence deterministically generates 128 candidates within ±0.05 m translation and 15° rotation, and simulation runs on a 7-DoF Franka Emika Panda in PyBullet.
The fuzzy rough lower approximation measures consistent support from strongly related neighbors, the upper approximation retains broader local support, and their difference forms the boundary region. Decisions are routed by region: candidates above the lower-approximation threshold τpos = 0.80 go to normal execution, otherwise boundary candidates meeting upper-support conditions go to compliant execution, and if neither exists the attempt is rejected. In the threshold-response analysis, raising τpos from 0.80 to 0.90 lowered the safety-violation rate from 2.1% to 1.4% but raised the rejection rate from 4.5% to 12.5% and cut success to 80.7%.
Perspective
The evaluation covers 1500 randomized tabletop scene-instruction instances with a 7-DoF Franka Emika Panda in PyBullet, plus 300 real executions on a Realman RM65-B with an Intel RealSense D435i depth camera; the real-robot part tests integration of the complete VLA-FRS-Grounding-control pipeline rather than a hardware-level comparison against each baseline.
Results are limited to the tested manipulation tasks, perturbation settings, robot platforms and system configurations, and the comparison characterizes the complete pipeline rather than the independent contribution of each internal component; the pairwise relation requires O(N²) evaluations and may become costly as the candidate set grows.