Research map / Reinforcement learning / Visual and representation RL
ABNet: Adaptive explicit-Barrier Net for Safe and Scalable Robot Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ABNet, a multi-head neural controller built from closed-form explicit-barrier (CBF-based) solutions whose heads are fused with trainable weights, giving provable safety, efficient training, and noise robustness across robot tasks.
Problem
Safe learning for robots is hard to scale: barrier-based differentiable QP layers are slow to train, give poor solutions, cannot be merged into larger models, and produce unstable controls under noisy inputs.
Approach
Derive an explicit closed-form solution of the differentiable CBF QP (for two aggregated barrier functions) so barriers appear explicitly as gates, then fuse multiple such heads (and other safe models like BarrierNet or dMPC) via convex trainable weights summing to one, which preserves forward invariance and smooths outputs.
Key results
- In vision-based driving (VISTA), ABNet variants achieve 0% crash and 100% pass rate, versus 2-4% crash and 33-39% pass for BarrierNet-related and DFB models.
- ABNet inference takes 0.004 s versus 0.872 s for the MPC baseline in driving.
- In 2D obstacle avoidance under 10% input noise, ABNet-100 reduces u2 control uncertainty to 0.013 versus 0.047 for BarrierNet while keeping theoretical guarantees.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ABNet instead of BNet for 2D obstacle avoidance | 2D robot obstacle avoidance · 2D robot obstacle avoidance | MSE: 0.014±0.006 -> 0.008±0.005 | No clear effect |
| Increase the one-shot ABNet head count from 10 to 100 (ablation) | ABNet · 2D robot obstacle avoidance | u2 uncertainty: 0.030 -> 0.013 | Helped |
| Use scalable training with 10 heads instead of one-shot training with 10 heads (ablation) | ABNet · Robot manipulation | MSE: 1.2e-4±9.6e-5 -> 5.9e-5±5.5e-5 | Helped |
| Increase the one-shot ABNet head count from 10 to 100 (ablation) | ABNet · Robot manipulation | u1 uncertainty: 0.008 -> 0.005 | Helped |
| Increase ABNet head count through scalable training (ablation) | ABNet · 2D robot obstacle avoidance | control uncertainty | Helped |
| Increase ABNet head count through scalable training (ablation) | ABNet · Robot manipulation | control uncertainty | Helped |
| Add attention images and use separate image regions for heads (ablation) | ABNet · Vision-based autonomous driving | SAFETY: 1.455 -> 4.198 | Helped |
| Add attention images and use separate image regions for heads (ablation) | ABNet · Vision-based autonomous driving | CRASH: 0% -> 0% | No clear effect |
| Scale training by combining ABNet with ABNET-ATT (ablation) | ABNet · Vision-based autonomous driving | CRASH: 0% -> 0% | No clear effect |
| Scale training by combining ABNet with ABNET-ATT (ablation) | ABNet · Vision-based autonomous driving | SAFETY: 1.455 -> 2.221 | Helped |
| Merge safe policies with ABNet rather than use a single BNet | 2D robot obstacle avoidance · 2D robot obstacle avoidance | u2 uncertainty: 0.047 -> 0.030 | Helped |
| Use ABNet rather than BNet | Vision-based autonomous driving · Vision-based autonomous driving | PASS: 33% -> 100% | Helped |
Surprising findings
- Existing model merging approaches such as E2E ensembles can achieve the best MSE but violate safety constraints (negative safety values).
- BarrierNet with dQP layers gives very poor training and testing solutions, degrading model quality.
Limitations
- All ABNet heads share the same safety constraints.
- Safety specifications must be known a priori; learning them from data is left for future work.
- Model merging is done in output space rather than parameter space, so parameter-level safety merging is not addressed.
Experimental setup
Datasets: VISTA simulator benchmark (vision-based driving) 2D robot obstacle avoidance simulation two-link planar manipulator simulation
Benchmarks: 2D robot obstacle avoidance closed-loop test two-link robot manipulation closed-loop test VISTA vision-based driving closed-loop test
Metrics: safety constraint satisfaction conservativeness MSE control uncertainty obstacle crash rate obstacle pass rate computation time
Baselines: E2E (Levine et al., 2016) V-E2E (Amini et al., 2022) BarrierNet DFB BNet-UP E2Es-MCD E2Es-DR dQP
Builds on: BarrierNet OptNet / differentiable QP high-order control barrier functions Deep forward-backward (DFB) Monte-Carlo dropout
Closest papers by meaning
- Verified Safe Reinforcement Learning for Neural Network Dynamic Models (NeurIPS 2024 · similarity 0.86)
- Safe Planner: Empowering Safety Awareness in Large Pre-Trained Models for Robot Task Planning (AAAI 2025 · similarity 0.86)
- A Physics-Informed Machine Learning Framework for Safe and Optimal Control of Autonomous Systems (ICML 2025 · similarity 0.85)
- SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learning (NeurIPS 2025 · similarity 0.84)
- Latent Safety-Constrained Policy Approach for Safe Offline Reinforcement Learning (ICLR 2025 · similarity 0.84)
- Boundary-to-Region Supervision for Offline Safe Reinforcement Learning (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.