Can a small robot planner learn formal safety rules without losing the verifier?
Post-training lifted compact models to 70.5%–76.0% success on symbolic planning tests, while a verifier blocked rejected plans from a desktop arm. The physical test covered only 30 trials in one block-stacking workcell.
The short version
- What happened
- Training used four symbolic robotics domains. The SFT and reinforcement-learning sets each held 1,000 prompts from disjoint sets of 500 planning problems; evaluation used a separate 200-instance test set.
- Who it affects
- Workers around robots could benefit from auditable high-level checks and local processing, but encoded rules must match the real workplace and motion-level safeguards must remain independent.
- What remains unproven
- The unreviewed paper evaluates authored symbolic constraints, adapted simulators and one desktop block-stacking workcell.
Keep in mind
The unreviewed paper evaluates authored symbolic constraints, adapted simulators and one desktop block-stacking workcell.
The physical experiment used 30 trials; verifier acceptance covered high-level plans, not perception, manipulation precision or unmodelled risks.
Energy figures are short device-level measurements, not comparisons of training energy or full-system lifecycle impact.
The paper reports no separate funding or competing-interest statement.
Source analysis · 1 linked source. Primary analysed source date: 20 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
Formal feedback trained the planner, then remained at runtime
SLoSa begins with four symbolic planning domains—Blocksworld, Ferry, Grippers and Spanner—and adds constraints such as required ordering, safe intermediate states, restricted resources or mandatory inspections. Classical planners and a formal verifier keep examples only when the original plan violates the added constraint and a valid alternative exists. The supervised and reinforcement-learning stages each used 1,000 prompts derived from disjoint sets of 500 underlying planning problems, with separate validation and test problems.
Three quantised models, with 4 billion, 8 billion and 14 billion parameters, were adapted with verified plans and then refined using verifier-labelled failures. On the common 200-instance test set, pretrained models scored zero because they did not follow the required action vocabulary and format. Supervised fine-tuning reached 59% to 67%; verifier-guided reinforcement learning lifted success to 70.5%, 74.0% and 76.0%. Separate adapted household simulators tested whether the pipeline travelled beyond PDDL, though one used a fully observed, non-interactive version rather than the standard protocol.[1]
The physical arm exposed the boundary between plan and motion
The researchers also measured local inference. Four-bit models on an Nvidia Jetson Orin averaged 17.85 to 26.41 watts, depending on model size. A verifier-and-repair loop raised the 8B model's final symbolic success to 77.5% while reducing average generation from 5,186 tokens for the untrained model to 1,783 after full post-training. These are device and task measurements, not a lifecycle energy comparison.
For the desktop arm, images and text were converted into a symbolic block-stacking problem. Only verifier-accepted plans could execute. Across 30 trials, the full model produced 19 accepted plans and completed 14; five accepted plans still failed because block placement lacked precision or the stack exceeded the arm's reachable workspace. That distinction is central: formal checks covered encoded high-level rules, not perception errors, motion control or unmodelled hazards. Broader robots, workcells, tasks and independent safety cases would be needed before deployment claims.[1]
The distinction that matters
Two safety layers, two different jobs
Training makes valid plans more likely; runtime verification decides whether a plan may execute.
Teach
Fine-tune on formally verified alternatives
Improves first-pass planning
Check
Verify each proposed action sequence
Blocks encoded constraint violations
Execute
Map an accepted plan to robot skills
Still faces perception and motion failures