Weekly Research Digest — 2026-07-26

22 new entries this week across 4 topic areas. This week’s cross-cutting sweep for VLA post-training / data-augmentation content turned up an unusually large batch — 17 of the 22 new entries are flagged high priority.


Vision-Language-Action (VLA) Models

ReleaseVenueSignificance
robottt-context-scaling-robot-policies RoboTTT: Context Scaling for Robot PoliciesarXiv⭐ HIGH PRIORITY: Test-time-training “fast weights” scale VLA context to 8K timesteps, enabling one-shot in-context imitation and on-the-fly policy improvement without extra inference latency.
flowdagger-human-in-the-loop-latent-space-policy-adaptation FlowDAgger: Human-in-the-Loop Adaptation of Generative Robot Policies in Latent SpacearXiv⭐ HIGH PRIORITY: “Action inversion” lets frozen flow/diffusion policies be DAgger-adapted from a handful of human corrections without full fine-tuning or online RL.
z1-efficient-rl-vla Z-1: Efficient Reinforcement Learning for Vision-Language-Action ModelsarXiv⭐ HIGH PRIORITY: GRPO-based RL post-training lifts π0.5 from 67.4% to 80.6% success across 24 RoboCasa tasks.
force-efficient-vla-rl-value-calibrated-warmup FORCE: Efficient VLA Reinforcement Fine-Tuning via Value-Calibrated Warm-up and Self-DistillationarXiv⭐ HIGH PRIORITY: A Q-function warm-up phase fixes catastrophic initial unlearning in VLA RL fine-tuning, +79% absolute success over imitation baselines.
ttt-vla-test-time-latent-prompt-optimization TTT-VLA: Test-Time Latent Prompt Optimization for Vision-Language-Action ModelsarXiv⭐ HIGH PRIORITY: Adapts a learned latent prompt (not model weights) online at deployment to handle distribution shift.
retrieve-dont-retrain-vla-new-tasks-test-time Retrieve, Don’t Retrain: Extending Vision Language Action Models to New Tasks at Test TimearXiv⭐ HIGH PRIORITY: A frozen retrieval-augmented policy adds new tasks by indexing demonstrations instead of fine-tuning.
expo-ft-sample-efficient-rl-finetuning-vla EXPO-FT: Sample-Efficient Reinforcement Learning Finetuning for Vision-Language-Action ModelsarXiv⭐ HIGH PRIORITY: Q-guided candidate selection plus human-in-the-loop corrections make RL fine-tuning of VLAs sample-efficient on precision/dynamic tasks.
supervise-what-survives-geometry-guided-vla-adaptation-synthetic-video Supervise What Survives: Geometry-Guided VLA Adaptation from Synthetic Robot VideosarXiv⭐ HIGH PRIORITY: Uses synthetic robot video only for geometric auxiliary supervision (not pseudo-actions), avoiding the sim-to-real action mismatch of prior approaches.
rynnbrain-11-embodied-foundation-model RynnBrain 1.1: Towards More Capable and Generalizable Embodied Foundation ModelarXivAlibaba DAMO’s 2B/9B/122B-A10B embodied foundation model family adds contact-point prediction and native 3D grounding, evaluated across three distinct robot embodiments.

World Models for Robotics

ReleaseVenueSignificance
temporal-ratio-video-action-generalization-gap Understanding and Mitigating the Video-Action Generalization Gap via Temporal RatioarXiv⭐ HIGH PRIORITY: Identifies why video-pretrained foundation models lose compositional generalization after being fine-tuned into VLAs, and fixes it with an inference-time guidance metric.
world-labs-acquires-scenix World Labs Acquires SceniXblog⭐ HIGH PRIORITY: Fei-Fei Li’s World Labs acquires a Stanford robotics-sim spinout to fuse generative 3D world models with simulation pipelines for synthetic robot training data.
gigaworld-1-world-models-robot-policy-evaluation GigaWorld-1: A Roadmap to Build World Models for Robot Policy EvaluationarXivOpen-sourced WMBench (324k+ rollouts, 7 world-model families) shows evaluator quality tracks action-faithfulness over long horizons, not photorealism.
mu0-scalable-3d-interaction-trace-world-model μ0: A Scalable 3D Interaction-Trace World ModelarXivPredicts embodiment-agnostic 3D interaction traces instead of pixels or actions, letting it pretrain from action-free video and transfer across embodiments.

Reinforcement Learning for Robotics

ReleaseVenueSignificance
freeform-preference-learning-robotic-manipulation Freeform Preference Learning for Robotic ManipulationarXiv⭐ HIGH PRIORITY: Multi-axis natural-language preferences (speed, safety, placement quality) give a denser reward signal than binary preference learning, +38pp over baselines.
trust-your-instincts-confidence-driven-test-time-rl-vla Trust Your Instincts: Confidence-Driven Test-Time RL for Vision-Language-Action ModelsarXiv⭐ HIGH PRIORITY: Test-time RL that uses a VLA’s own generation confidence as an intrinsic reward, approaching oracle RL performance with no ground-truth reward.
leveraging-offline-supervision-rl-large-scale-vla Leveraging Offline Supervision for Efficient and Generalizable Reinforcement Learning in Large-Scale Vision-Language-Action ModelsarXiv⭐ HIGH PRIORITY: Hybrid offline-online RL training retains RL’s OOD generalization advantage while cutting the costly online interaction needed to get there.
sarl-semantic-reinforcement-learning-generalist-robot-policies Adapting Generalist Robot Policies with Semantic Reinforcement Learning (SARL)arXiv⭐ HIGH PRIORITY: Adapts generalist VLAs by running RL over language prompts instead of raw actions, composing existing skills for novel long-horizon OOD tasks.
enpire-agentic-robot-policy-self-improvement-real-world ENPIRE: Agentic Robot Policy Self-Improvement in the Real WorldarXiv⭐ HIGH PRIORITY: AI coding agents run the full policy-improvement loop on a real robot fleet, hitting 99% success on dexterous tasks and halving training time from 1→8 robots.

Humanoid Robotics

ReleaseVenueSignificance
legs-teleop-free-vla-humanoid-gaussian-splatting-world LEGS: Fine-Tuning Teleop-Free VLAs for Humanoid Loco-manipulation in an Embodied Gaussian Splatting WorldarXiv⭐ HIGH PRIORITY: Gaussian-splatting simulation generates teleop-free fine-tuning data for humanoid VLAs that matches or beats real teleoperated demonstrations.
humanoidumi-robot-free-demonstrations-humanoid-whole-body-manipulation HumanoidUMI: Bridging Robot-Free Demonstrations and Humanoid Whole-Body ManipulationarXiv⭐ HIGH PRIORITY: Portable VR/handheld gripper rig collects humanoid whole-body manipulation demonstrations with no robot or teleoperation hardware required.
1x-neo-hands-tendon-driven-dexterous-hand 1X NEO Hands — 25-DoF tendon-driven, force-transparent dexterous handblog1X’s new NEO hand uses quasi-direct-drive tendons for native force sensing/control, demoed on LEGO assembly, screwdriver use, and sign language.
boston-dynamics-atlas-training-humanoid-hard-work Training a Humanoid Robot for Hard WorkblogGPU-parallel simulation RL lets Atlas lift 100+ lb loads using proprioceptive/force sensing rather than vision, generalizing zero-shot beyond its training weight range.

Generated automatically. All entries verified via web search.