A generalizable safety-aware robot behavior planning method based on an explicit constraint space
A Generalizable Approach to Safety-Aware Robotic Behavior Planning with an Explicit Constraint Space
The authors propose a safe robot behavior planning framework with an explicit Constraint Space, and the authors say it achieves a constraint-selection F1 of 0.3881 on 150 held-out queries, outperforming BGE-M3 retrieval.
This is not an attack paper: the goal is for the robot to satisfy safety constraints in tabletop manipulation that vary with task and scene and are not explicitly written in the instruction. Before execution, the Safety Inspector checks the plan, and three failed checks invoke call_human_help(); no adversarial attacker is involved.
- In desktop robot manipulation, safety constraints vary with task and scene and are often not written in the instruction; the authors argue that retrieving rules by semantic similarity alone cannot reflect whether a constraint truly applies in the current scene.
- Maintains an explicit Constraint Space, uses a query-conditioned two-tower model to learn constraint applicability, has a multimodal Safety Inspector check plans and reconstruct the complete constraint set, and uses verified feedback for test-time adaptation (including LoRA) and constraint library expansion.
- On 150 held-out queries, constraint selection F1 is 0.3881, higher than 0.2519 for BGE-M3 retrieval (Table I); the Inspector with Qwen2.5-VL-72B+LoRA achieves 97.7% judgment accuracy and 67.1% exact correction match rate (Table II); there are also two real robot examples.
- The authors note that the framework operates only at the behavior planning layer and does not replace underlying obstacle avoidance, motion-level safety control, or formal verification; the evaluation is limited to desktop manipulation and textual safety constraints, and the stability of continuous adaptation under longer deployment sequences has not been systematically evaluated (Sec. V).
- Evaluated models
- Qwen3-VL-30B-A3BQwen2.5-VL-32BQwen2.5-VL-72B
- Benchmarks
- 自建桌面场景基准(50 个原始场景 + 50 个新场景,150 个留出任务-场景查询,统一 Constraint Space 含 93 条规则)30 个未见场景(390 个检查样本)
- Metrics
- RecallPrecisionF1EMRIDACEMRECHVPR
The study proposes a safety-aware robot behavior planning framework that uses an explicit constraint space to store reusable textual safety rules. A lightweight dual-tower model selects applicable constraints according to task and environment for use by a multimodal behavior planner; a Safety Inspector checks primitive plans, reconstructs the complete constraint set, and provides corrective feedback. Verified feedback is used for test-time adaptation of constraint applicability and for LoRA fine-tuning of the Safety Inspector, and new constraints are incorporated into the constraint space.
Paper analysis
What problem does this paper attempt to solve?
Addresses the problem in desktop robot planning of safety constraints varying with task and scene, and of constraint selection and execution safety when constraints are not explicitly written in the instruction.
Core problem: safety constraints in tabletop manipulation depend on task and scene and are often not explicitly written in the instruction; how can the robot explicitly select and satisfy these constraints during planning.
- Scenario and importance: the authors believe LLM/VLM can already generate long-horizon action sequences from natural language instructions and visual observations, but completing a task is not the same as completing it safely; the same instruction has different safety requirements in different scenes (e.g., handling liquids near electronic devices).
- Shortcomings of existing methods: the authors believe leaving safety knowledge entirely to the planner's implicit reasoning is unreliable; if a reusable rule library is maintained, retrieving rules only by semantic similarity cannot reflect whether a constraint should actually be enforced in the current scene.
- Core observation: the key problem is to jointly determine the applicability of each constraint conditioned on task and environment; and a fixed rule library and a fixed applicability model cannot cover all safety situations encountered at test time.
- Paper's approach: proposes a safety-aware behavior planning framework based on an explicit Constraint Space: a query-conditioned dual-tower model selects applicable constraints, a multimodal Safety Inspector checks the plan and reconstructs constraints, and verified feedback is used for test-time adaptation and constraint library expansion.
- Threat model:
- The paper is not an attack study and has no adversarial attacker; the source of risk is scene-relevant safety hazards and unsafe planner outputs.
- Protected object: tabletop robot behavior planning driven by a natural language task Tt and an environment image It.
- Defense position: pre-execution checking, with only PASS plans executed; three failed checks invoke call_human_help().
- The paper does not discuss adversarial scenarios such as prompt injection or malicious instructions.
What related research exists?
Related work falls into three groups: safety-aware robot manipulation, LLM task planning, and test-time adaptation; the baseline is BGE-M3 semantic retrieval.
Safety-aware robot manipulation
- Semantically Safe Robot Manipulation [5]: combines semantic scene understanding with language model reasoning to identify unsafe conditions at the semantic level and map them to motion-level safeguards.
- Safety-as-Policy [6]: requires the robot to consider potential hazards and their consequences while completing manipulation instructions, letting safety considerations influence behavior generation.
- Safety Chip [7]: converts natural language safety constraints into Linear Temporal Logic specifications, monitors actions generated by the LLM, and blocks execution of unsafe actions. The paper notes the problem this group of work leaves in common: when safety knowledge is maintained as a reusable set of constraints, how to determine which constraints apply to the current task and environment, and how the knowledge evolves with new situations.
LLM robot task planning
- ProgPrompt [8] / Code as Policies [9]: the former uses programmatic environment descriptions to let the LLM generate situated task plans, while the latter treats the language model as a policy that writes code, generating robot programs that compose perception and control APIs.
- SayCan [10]: combines the semantic reasoning of language models with the affordance of pretrained skills to select actions that are both relevant and executable. The paper notes that task relevance and physical feasibility alone cannot guarantee safe behavior in desktop environments.
Test-time adaptation
- Test-time training [11]: uses self-supervised signals at inference time to improve robustness under distribution shift.
- LoRA [12]: a parameter-efficient adaptation method that updates only a small number of additional parameters while freezing the backbone. The paper notes that this work differs from self-supervised test-time adaptation: the adaptation signal is generated by the closed-loop Safety Inspector and enters the system only after human or post-review verification.
Baselines and positioning
- Baseline method: BGE-M3 [13] semantic retrieval, taking the Top-nt rules by cosine similarity for each query (nt is the same number as selected by our model), compared on a unified Constraint Space containing 93 rules (Table I).
- The authors position this work as: formalizing constraint selection as task-environment-conditioned applicability learning, and letting test-time improvement occur simultaneously in constraint selection, safety checking, and reusable knowledge representation.
How does the paper address this problem?
Explicit Constraint Space + query-conditioned dual-tower applicability model + multimodal Safety Inspector closed loop + test-time adaptation driven by verified feedback.
Overall approach: place safety knowledge in an explicit Constraint Space outside the planner, first perform task-environment-conditioned constraint selection, then have a Safety Inspector check and correct in a closed loop before execution, and feed verified feedback back for adaptation.
Problem setup and overall pipeline
- Given a task Tt and an image It, select an applicable subset St from the Constraint Space Ct, and the planner generates Pt = P(Tt, It, St).
- The plan is checked by the Inspector and executed only if PASS; otherwise it is replanned with the reconstructed constraint set Ĝt, looping until PASS.
- After three failed checks, call_human_help() is invoked as a safety fallback.
Query-conditioned constraint applicability learning
- The VLM converts the image into a textual description Dt = V(It), which is concatenated with the task into a query Qt = Tt ⊕ Dt; the frozen BGE-M3 encodes the query and each constraint (1024 dimensions).
- Two-tower structure: the rule tower fr is a 1024→1024→512 MLP; the query tower fq is a 1024→1024→513 MLP, predicting a separating hyperplane (wt, bt) for each query.
- Constraint applicability is determined by the normalized signed distance: dt,i = (xi⊤ wt + bt) / |wt|2, and at inference St = {ci | dt,i > 0}; sigmoid is used only for training.
- Offline training uses weighted binary cross-entropy with positive class weight αt = Nt−/Nt+ to handle class imbalance, optimizing only the two towers with the encoder frozen.
Multimodal Safety Inspector and constraint reconstruction
- The inspection state Xt = (Tt, It, St, Pt), and the output is PASS or (FAIL, Ĝt, P̂t): Ĝt is the reconstructed complete applicable constraint set (keeping applicable ones, removing inappropriate ones, adding missing ones), and P̂t is the corrected plan proposal.
- The reconstructed constraints (rather than the plan proposed by the Inspector) are sent back to the planner for replanning; P̂t is not executed directly and serves only as structured correction information.
Test-time adaptation and knowledge expansion
- Post-review verification: a prompt-constrained post-reviewer gives ρt ∈ {PASS, REJECT} for the correction result, and only PASS corrections enter adaptation, which the authors state is to reduce the risk of propagating erroneous self-generated feedback.
- Applicability model adaptation: with Ĝt as positive samples and St \ Ĝt as negative samples, the two towers are updated with standard binary cross-entropy; constraints in neither St nor Ĝt remain unlabeled.
- Inspector adaptation: updated with LoRA with the VLM backbone frozen; correction information found in later rounds is associated with the initial failure state to construct single-round correction training samples.
- Constraint library expansion: ΔCt = Ĝt \ Ct is merged into the Constraint Space for use in subsequent planning.
Implementation and primitive grounding
- Plans are represented as sequences of predefined operation primitives; Grounding DINO [14] localizes targets, SAM2 [15] generates masks, and CoPa [16] generates spatial constraints.
- GPT-5.5 [17] is used for environment description generation, multimodal behavior planning, and post-review; the Inspector backbone is configured separately per experiment.
What experiments did the paper conduct?
Four sets of experiments: constraint applicability (150 queries), Inspector comparison of three backbones (390 samples), sequential test-time adaptation over 50 scenes, and two real robot examples.
Experimental Setup
- Subjects under test: constraint applicability dual-tower model (BGE-M3 encoder frozen); Safety Inspector with three VLM backbones: Qwen3-VL-30B-A3B, Qwen2.5-VL-32B, Qwen2.5-VL-72B, all with LoRA added and backbone frozen.
- Benchmark: self-built tabletop scenes (home, kitchen, office, laboratory); applicability evaluation uses 50 original scenes for training and 50 new scenes for testing, each scene paired with 3 task configurations, totaling 150 held-out task-scene queries, with a unified Constraint Space containing 93 candidate rules; Inspector evaluation uses 30 unseen scenes totaling 390 inspection samples.
- Baseline: BGE-M3 Top-nt semantic retrieval, returning the same number of rules per query as this paper's model.
- Metrics: applicability uses Recall, Precision, F1, EMR (the paper treats Recall as the primary safety metric); Inspector uses IDA (PASS/FAIL decision accuracy), CEM (among samples correctly judged FAIL, the proportion whose corrected output exactly matches), REC (missing-rule reconstruction coverage q/(m−n)); the adaptation experiment additionally uses HVPR (the proportion of scenes where the Inspector PASSes and manual verification confirms safety and task completion).
- Evaluation method: applicability and Inspector metrics are compared against manually annotated ground truth; HVPR includes manual review; post-review is given PASS/REJECT by a prompt-constrained post-reviewer.
- Sample sizes: 150 queries; 390 inspection samples; 50 scenes split into 5 stages, 10 per stage; the paper does not report the number of repetitions.
Main Results
Constraint applicability (Table I, 150 held-out task-scene queries, both methods return the same number of rules per query):
Swipe horizontally to read the table
Method Recall ↑ Precision ↑ F1 ↑ EMR ↑ BGE-M3 Top-nt 24.32% 26.11% 0.2519 0.00% Ours 37.48% 40.23% 0.3881 2.67% - The authors state that under the same rule-selection budget, this paper's model outperforms semantic retrieval on all four metrics.
- The authors also note that absolute Recall under distribution shift remains limited.
- The authors believe the improvement shows that the learned model can capture task-environment-relevant constraint applicability beyond semantic similarity.
Safety Inspector Evaluation
- What was tested: IDA/CEM/REC of three LoRA-adapted VLM backbones on 30 unseen scenes and 390 inspection samples (Table II).
- Results: IDA is 90.0% (Qwen3-VL-30B-A3B), 96.7% (Qwen2.5-VL-32B), 97.7% (Qwen2.5-VL-72B); CEM is 35.3%, 49.3%, 67.1%; REC is 97.8%, 97.8%, 98.1%.
- Authors' interpretation: differences among backbones are most evident on CEM, while REC remains stable across configurations; the authors believe generating a complete corrected output is harder than making the initial PASS/FAIL decision.
Sequential Test-Time Adaptation
- What was tested: 50 original scenes are processed sequentially in 5 stages (10 per stage); each stage is first evaluated and then the verified feedback is used for three adaptation mechanisms; before Stage 0 the model was not trained or adapted on these scenes (Table III).
- Results: Recall goes from 0.00% at Stage 0 to 46.15% at Stage 2 and 23.08% at Stage 4; Precision goes from 0.00% to 60.00% at Stage 4; HVPR goes from 30.00% to 50.00% at Stage 4, with intermediate Stages 2 and 3 both at 20.00%.
- Authors' interpretation: sequential feedback moves the model out of its initial zero-selection state and improves constraint prediction over deployment, but because scene difficulty differs across stages, performance is not strictly monotonic; HVPR in the final stage is higher than cold start.
- Exception note: the authors explicitly state that Inspector LoRA adaptation and constraint library expansion were enabled throughout, but Inspector-specific metrics were not recorded per stage, so the results characterize the behavior of the full adaptation framework rather than the separate contribution of Inspector adaptation.
Real Robot Experiments
- What was tested: two representative tasks on a physical platform (put the ball into the pen holder; put the tape onto the shelf), comparing direct plans with safe plans after adding scene-relevant constraints (Figure 2).
- Results: the safe plan first uses Grasp/Put primitives to move away nearby bottles or scissors before executing the original task; the paper does not report quantitative metrics such as success rate.
- Authors' interpretation: the framework modifies the primitive sequence according to scene safety requirements while preserving the original task goal, and the inserted safety operations reuse predefined primitives, making it compatible with existing deployment execution pipelines.
Other Ablations and Analyses
- The paper does not report separate ablation experiments; the contribution of Inspector adaptation is not isolated (see above).
What are some points that can be further explored?
The authors list three types of limitations—planning-layer positioning, evaluation scope, and long-term adaptation stability—and propose future directions such as temporal constraints.
Limitations and future directions noted by the authors
- Does not replace underlying safety mechanisms: the framework operates at the behavior planning layer and does not replace underlying collision avoidance, motion-level safety control, or formal verification (Sec. V).
- Evaluation scope: the current evaluation is limited to desktop manipulation and textual safety constraints (Sec. V).
- Long-term stability not evaluated: the long-term stability of continuous Inspector adaptation under longer deployment sequences has not been systematically evaluated (Sec. V).
- Future directions: the authors state that future work will consider temporal constraints, richer physical interaction, longer-horizon manipulation tasks, and longer-term adaptation (Sec. V).
- Granularity of metric recording: in the sequential adaptation experiments, Inspector-specific metrics were not recorded for each stage, and the results reflect the complete framework rather than the separate contribution of Inspector adaptation (Sec. IV-C and Sec. V).
Scope of experiments
- Models tested: the Inspector backbones are Qwen3-VL-30B-A3B, Qwen2.5-VL-32B, and Qwen2.5-VL-72B, all with LoRA; planning, environment description, and post-review use GPT-5.5.
- The applicability evaluation covers 150 held-out task-scene queries and 93 candidate rules; the Inspector evaluation covers 30 unseen scenes and 390 inspection samples.
- The sequential adaptation experiment covers 50 scenes and 5 stages, with 10 scenes per stage.
- The real robot experiments are qualitative examples of 2 representative tasks, and the paper does not report quantitative success rates.
- The paper does not report the number of experimental repetitions, the scale of training data for Inspector adaptation, or the agreement between post-reviewer verification and human judgment.
Summarize the main content of the paper
A safe robot behavior planning framework with an explicit constraint library + applicability learning + closed-loop safety checking + test-time adaptation.
A paper in the direction of robot safety that proposes a safety-aware behavior planning framework centered on an explicit Constraint Space, addressing the problem in tabletop manipulation where safety constraints vary with task and scene and are not explicitly written in the instruction.
- Method: a query-conditioned dual-tower model learns constraint applicability conditioned on task and environment (at inference, selection is by the sign of the signed distance); a multimodal Safety Inspector checks the plan before execution, reconstructs the complete constraint set and gives correction proposals, and execution proceeds only on PASS, with three failed checks invoking human fallback; feedback verified by the post-reviewer is used for applicability model updates, LoRA adaptation of the Inspector, and constraint library expansion.
- Constraint selection: on 150 held-out task-scene queries, this paper's model achieves Recall 37.48% and F1 0.3881, higher than BGE-M3 retrieval's 24.32% and 0.2519 under the same rule-count budget (Table I); the authors also note that absolute Recall under distribution shift remains limited.
- Inspector: Qwen2.5-VL-72B+LoRA achieves IDA 97.7%, CEM 67.1%, REC 98.1% on 390 inspection samples (Table II); the authors believe complete correction generation is harder than PASS/FAIL decision.
- Test-time adaptation: in the 50-scene, 5-stage sequential experiment, HVPR goes from 30.00% at Stage 0 to 50.00% at Stage 4, but stages are non-monotonic, and the separate contribution of Inspector adaptation is not isolated (Table III).
- Real robot: two qualitative examples show that the safe plan first moves away nearby hazards or obstructions before executing the original task (Figure 2); the paper gives no quantitative results.
- Authors' conclusion: the experiments show that combining explicit safety knowledge with deployment-time correction feedback is effective; the authors also explicitly state that the framework does not replace low-level safety control and formal verification, and long-term adaptation stability is left for future work.