#VLA
6 related articles
Backtrack: A Lightweight Failure Recovery Mechanism Makes Frozen VLA Robots More Robust
Backtrack is a lightweight failure detection and recovery mechanism for frozen VLA models, using a small prediction head to detect failures and backtrack for retries—cutting failure rates by 19-48% on LIBERO-Long without retraining.
AtomicRL: Atomic Reinforcement Learning Solves Long-Horizon Robotic Manipulation Challenges
AtomicRL introduces atomic reinforcement learning with Semantic-Aware Policy Optimization (SAPO) and Bottleneck-Adaptive Targeted Exploration (BATE) to solve credit assignment and subtask imbalance in VLA long-horizon robotic manipulation, outperforming PPO on LIBERO-LONG and CALVIN and improving real-world task performance by 21.3%.
PaceVLA: Adaptive Temporal Action Quantization for Faster Robot Execution
PaceVLA proposes an adaptive temporal action quantization framework that boosts VLA model execution efficiency via pattern-specific action decoders and quantization-guided routing, validated in simulation and real-world experiments.
Integrating World Models into VLAs: A Comparison of Four Paradigms and Key Findings
A systematic study categorizes VLA world model integration into four paradigms and finds under a unified evaluation framework that 2D semantic representations are robust across paradigms, decoupled MoT Future Experts excel in OOD robustness, and world modeling yields the largest gains in long-horizon and dynamic tasks.
Robot Execution Horizon Is Not a Tuning Knob: A Staleness–Amortization Boundary Law
The robot VLA execution horizon is no tuning knob but the solution to a stationarity condition. This study derives the objective, proves the horizon is set by inference price and plant latency-cost curvature, exposes a two-ended feasible window and a latency wall, and refutes the square-root law.
Splat2Act: Scan Your Home with a Phone to Teach Robots How to Grab Your Stuff
Splat2Act is a real-to-sim-to-real pipeline that lets users phone-scan their environment to personalize VLA policies like π0.5, boosting grasp success from 0% to 63.9%.