P01 Vision Language Action Models in Robotic Manipulation: A Systematic Review
A model–data–simulation map, dataset description scores, and checks on heterogeneous evaluation claims.
LITERATURE NOTES
Reading
Research questions, core methods, experimental evidence, limitations, and connections between papers.
P02 A Survey on Vision-Language-Action Models: An Action Tokenization Perspective
Eight action-related interfaces connect modules, data sources, and execution, with overlapping categories.
P03 Survey of Vision-Language-Action Models for Embodied Manipulation
Architecture, data, pretraining, post-training, and evaluation, with corrections to surveyed descriptions.
P04 A Survey on Reinforcement Learning of Vision-Language-Action Models for Robotic Manipulation
Action likelihoods, rewards, transitions, learning paradigms, and the practical cost of RL-VLA deployment.
P05 RL²: Fast Reinforcement Learning via Slow Reinforcement Learning
A recurrent policy learns to explore and adapt through hidden state across episodes within a task.
P06 SimBa: Simplicity Bias for Scaling Up Parameters in Deep Reinforcement Learning
Running normalization and residual networks improve RL scaling; evidence and proxy limitations are examined.
P07 Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware
ALOHA hardware, CVAE action chunks, and temporal ensembling for fine bimanual imitation learning.
P08 Diffusion Policy: Visuomotor Policy Learning via Action Diffusion
Conditional action diffusion, receding-horizon control, and the distinction between sampling and robot time.
P09 π₀: A Vision-Language-Action Flow Model for General Robot Control
VLM semantics and a continuous flow action expert, with explicit data, timing, and version boundaries.
P10 MetaVLA: Unified Meta Co-training For Efficient Embodied Adaption
Demonstration context and deterministic meta co-training improve unified multitask VLA adaptation.
P11 π_RL: Online RL Fine-tuning for Flow-based Vision-Language-Action Models
Flow-Noise and Flow-SDE connect flow action generation to PPO-style online adaptation.
P12 RL Token: Bootstrapping Online RL with Vision-Language-Action Models
A frozen VLA supplies compact state and reference actions to a small online actor–critic.
P13 What Can RL Bring to VLA Generalization? An Empirical Study
Visual, semantic, and execution OOD reveal where PPO improves VLA behavior and where it does not.
P14 Action-aware Dynamic Pruning for Efficient Vision-Language-Action Manipulation
Action-aware visual-token pruning reduces computation, with limits of action magnitude as a dynamics proxy.
P15 Realtime-VLA FLASH: Speculative Inference Framework for Diffusion-based VLAs
Speculative drafts, parallel verification, and cache refresh trade fresh feedback against verification cost.
P16 Real-Time Execution of Action Chunking Flow Policies
Prefix-guided flow inpainting overlaps inference and execution while preserving chunk continuity.
P17 Running VLAs at Real-time Speed
CUDA Graphs and operator changes accelerate π₀; the measured implementation is separated from streaming proposals.
P18 VLASH: Real-Time VLAs via Future-State-Aware Asynchronous Inference
Future-proprioception conditioning supports asynchronous execution, without predicting future object states.
P19 DynamicVLA: A Vision-Language-Action Model for Dynamic Object Manipulation
A compact multiframe policy, dynamic data, and stale-prefix removal target moving-object manipulation.
P20 Towards Generalizable Robotic Manipulation in Dynamic Environments
DOMINO data and PUMA motion representations study dynamic manipulation under a bounded simulator protocol.
P21 MotionBricks: Scalable Real-Time Motions with Modular Latent Generative Model and Smart Primitives
Modular motion latents and smart primitives synthesize real-time kinematic references.
P22 Uni-Inter: Unifying 3D Human Motion Synthesis Across Diverse Interaction Contexts
Voxel-conditioned joint-heatmap diffusion unifies interaction synthesis across human, object, and scene contexts.
P23 GRAIL: Generating Humanoid Loco-Manipulation from 3D Assets and Video Priors
Known 3D assets and video priors produce physics-trackable humanoid loco-manipulation demonstrations.
P24 NavDP: Learning Sim-to-Real Navigation Diffusion Policy with Privileged Information Guidance
Privileged simulation supervision and candidate scoring train an RGB-D generative navigation policy.
P25 FLUX: Accelerating Cross-Embodiment Generative Navigation Policies via Rectified Flow and Static-to-Dynamic Learning
Rectified flow and dynamic reinforcement learning extend navigation, with careful social-metric definitions.
P26 NavIsaacLab: Generating Realistic Crowd via Parallel Robot Learning for Benchmarking Human-aware Navigation
Parallel crowd-path learning and full-body control form a human-aware navigation benchmark pipeline.
P27 Arena-Bench 2.0: A Comprehensive Benchmark of Social Navigation Approaches in Collaborative Environments
Nav2 integration and collaborative scenarios expose the need for shared success and event definitions.
P28 Social robot navigation: a review and benchmarking of learning-based methods
Learning-based social planners are reviewed and benchmarked, with implementation and budget caveats.
P29 Towards Physically Executable 3D Gaussian for Embodied Navigation
Object-aware Gaussian appearance is paired with existing mesh-derived collision geometry for navigation.
P30 Lyra 2.0: Explorable Generative 3D Worlds
Memory and camera-conditioned video generation build explorable worlds, with static-world and geometry limits.
P31 A Closer Look at Deep Policy Gradients
Gradient estimates, value prediction, and surrogate reward landscapes.
P32 Deep Reinforcement Learning that Matters: Supplementary Evidence
How seeds, architectures, reward scales, and implementations change algorithm comparisons.
P33 Era of Experience: Notes on Experience-Driven Learning
Separate experience-driven learning as a paradigm from world models as a mechanism.
P34 Fast Gradient-Descent Methods for Temporal-Difference Learning with Linear Function Approximation
MSPBE and auxiliary variables connect off-policy stability to linear-time updates.
P35 For SALE: State-Action Representation Learning for Deep Reinforcement Learning
SALE objectives, TD7 engineering components, and the scope of empirical evidence.
P36 High-Dimensional Continuous Control Using Generalized Advantage Estimation
TD-residual advantages, reward-to-go, lambda returns, and critic error.
P37 Policy Gradient Methods for Reinforcement Learning with Function Approximation
Occupancy distributions, compatible features, and theoretical assumptions for approximation.
P38 Potential-Based Shaping and Q-Value Initialization are Equivalent
A Q-value shift explains conditional path equivalence between shaping and initialization.
P39 Reinforcement Learning: An Introduction — A Reading Route
A chapter route from tabular methods through approximation, exploration, and policy gradients.
No published notes in this topic.