Weekly Research Digest — 2026-08-30

Weekly Research Digest — 2026-08-30

36 new entries this week across 4 topic areas, with a strong showing (16 entries) in the VLA post-training / data-scaling cross-cutting priority.

---

Vision-Language-Action (VLA) Models

Release	Venue	Significance
synthicl-scalable-in-context-imitation-learning-synthetic-data SynthICL: Scalable In-context Imitation Learning with Synthetic Data	arXiv	⭐ HIGH PRIORITY: RGB-only synthetic data trains in-context imitation policies hitting 79% success on 16 unseen real tasks from one demo.
e-tts-embodied-test-time-scaling-framework-robotic-manipulation E-TTS: A New Embodied Test-Time Scaling Framework for Robotic Manipulation	arXiv (ECCV 2026)	⭐ HIGH PRIORITY: history-aware iterative refinement with VLM verifiers gives up to +33% (sim) / +27% (real) on top of standard VLAs.
self-improving-vla-selected-diffusion-noise-spurious-robust-action Self-Improving VLA Policies: Selected Diffusion Noise	arXiv	⭐ HIGH PRIORITY: training-free test-time noise selection reduces spurious-cue reliance across π0, GR00T-N1.5/1.6.
logic-vla-temporal-logic-conditioned-vision-language-action-model Logic-VLA: A Temporal Logic Conditioned VLA Model	arXiv	⭐ HIGH PRIORITY: STL-conditioned VLA trained with flow-matching preference optimization over safety-satisfying vs. violating rollouts.
vane-reliable-test-time-training-vla-future-visual-representation VANE: Reliable Test-Time Training for VLA via Future Visual Representation Prediction	arXiv	⭐ HIGH PRIORITY: TTT that conditions updates on the future visual consequences of actions to avoid corrupting subsequent steps.
robosynchallenge-mastering-real-world-dexterity-synthesized-manipulation-skills RoboSynChallenge: Mastering Real-World Dexterity via Synthesized Manipulation Skills	arXiv (NeurIPS 2026 competition)	⭐ HIGH PRIORITY: competition benchmark explicitly built to test synthetic-data-trained bimanual policies against real-world-only scoring.
elastic-efficiently-learning-adaptively-scale-test-time-compute ELASTIC: Adaptively Scaling Test-Time Compute for Generative Control Policies	arXiv (CMU)	⭐ HIGH PRIORITY: meta-MDP allocates sequential vs. parallel test-time compute for diffusion/flow policies, cutting π0.5 latency 34% at matched success.
g05-galaxea-autoregressive-stream-robot-reasoning-action G0.5 (Galaxea): One Autoregressive Stream for Robot Reasoning and Action	arXiv	Major new open VLA foundation model — single decoder stream unifies reasoning and action tokens, no separate VLM encoder + action head.
ervla-revisiting-embodied-chain-of-thought-generalizable-robot-manipulation Revisiting Embodied Chain-of-Thought for Generalizable Robot Manipulation	arXiv	Largest embodied-CoT corpus to date (978K trajectories); reasoning-dropout training hits 86.9% on LIBERO-Plus.
robodojo-unified-sim-real-benchmark-generalist-robot-manipulation RoboDojo: A Unified Sim-and-Real Benchmark for Generalist Robot Manipulation Policies	arXiv	42 sim + 18 real tasks scoring generalization, memory, precision and long-horizon execution in one benchmark.
active-real-world-factor-based-evaluation-generalist-robot-policies Active Real-World Factor-Based Evaluation for Generalist Robot Policies	arXiv	Frames policy evaluation as sequential experimental design over task-factor space, validated against 700+ real evals per task.
univiewvla-unified-multiview-vision-language-action-world-modeling UniviewVLA: A Unified Multiview VLA Model with World Modeling	arXiv	Infers multiview future scene evolution from two cameras, lifting occlusion-task success from 40.0%→73.3%.
baton-long-horizon-robot-manipulation-agentic-subtask-exploration Don’t Drop the BATON: Long-Horizon Manipulation via Agentic Subtask Exploration	arXiv	Coding-agent layer drives long-horizon tasks over a frozen VLA via test-time exploration and checkable handoff contracts.
ucag-p-unified-camera-centric-action-geometry-pretraining One Policy, Many Embodiments: Unified Camera-Centric Action Geometry Pre-training	arXiv (Xiaomi)	Maps arms, humanoids and human hands into a shared camera-frame motion space, enabling joint robot+human training.
streampi-streaming-multimodal-temporal-modeling-vla StreamPI: Streaming Multimodal Temporal Modeling for VLA Models	arXiv (HKU-SAIL)	Adds streaming temporal reasoning to single-frame π0.5 with only ~9.2ms added latency for 5-frame context.
muvla-recurrent-memory-partially-observable-manipulation-vla μVLA: On Recurrent Memory for Partially Observable Manipulation in VLA Models	arXiv	Controlled ablation isolating what pure in-backbone recurrence buys over prior memory-VLA approaches.
revisiting-parameter-redundancy-vla-vlm-to-vla-adaptation Revisiting Parameter Redundancy in VLA Models	arXiv	Challenges the assumption that pruned VLA parameters are truly redundant during VLM-to-VLA adaptation.

World Models for Robotics

Release	Venue	Significance
wam-ttt-steering-world-action-models-watching-human-play-test-time WAM-TTT: Steering World-Action Models by Watching Human Play at Test Time	arXiv	⭐ HIGH PRIORITY: test-time training absorbs raw human video into a frozen WAM’s adaptive memory via self-supervised prediction.
zero-wam-in-context-world-action-modeling-human-videos Zero-WAM: In-Context World-Action Modeling from Human Videos	arXiv	⭐ HIGH PRIORITY: HumanGen pipeline synthesizes 74.2K human-robot ICL pairs; zero-shot task following from a human video prompt.
wall-ss-scaling-long-horizon-world-models-next-scale-autoregression WALL-SS: Scaling Long-horizon World Models via Next-Scale Autoregression	arXiv	Coarse-to-fine scale-wise generation plus scale-compressed memory targets coherent minute-long rollouts under bounded memory.
clap-cross-embodiment-video-world-models-zero-shot-physical-simulators CLAP: Cross-Embodiment Video World Models are Zero-Shot Physical Simulators	arXiv	Trains on internet-scale human + robot video toward a universal, embodiment-agnostic physical simulator.
phizero-world-model-physical-language PhiZero: A World Model Built Around Physical Language	arXiv (CASIA)	“Reason-then-render”: predicts a compact discrete physical-language sequence before diffusion-decoding to video, ~175x token reduction.
graphop-wm-world-models-morphology-parameter-generalization Graph-Operator World Models for Morphology-Parameter Generalization	arXiv	Factorizes dynamics into a morphology-independent basis plus a structured operator for generalizing across robot body variants.
maskwam-unifying-mask-prompting-prediction-world-action-models MaskWAM: Unifying Mask Prompting and Prediction for World-Action Models	arXiv	Object-centric mask prediction plus first-frame mask prompts cut language ambiguity in cluttered scenes.

Reinforcement Learning for Robotics

Release	Venue	Significance
r3-training-robots-reason-natural-language-reinforcement-learning R³: Training Robots to Reason in Natural Language via Reinforcement Learning	arXiv (CMU)	⭐ HIGH PRIORITY: mid-training on reasoning traces + rubric-based RL turns a VLM into a robotic reasoner guiding low-level control.
beyond-imitation-self-improving-robot-policies-off-policy-q-planning Beyond Imitation: Self-Improving Robot Policies via Off-Policy Q-Planning	arXiv (Georgia Tech)	⭐ HIGH PRIORITY: a small Q-function bolted onto a frozen BC policy enables online self-improvement without touching billion-parameter weights.
rynnvalue-scaling-robotic-value-foundation-models-temporal-distance RynnValue: Scaling Robotic Value Foundation Models with Temporal Distance	arXiv (Alibaba DAMO)	⭐ HIGH PRIORITY: label-free temporal-distance value model lifts real-world RL fine-tuning success from 52.5%→72.5%.
dipod-diffusion-policy-optimization-without-drifting-apart DiPOD: Diffusion Policy Optimization without Drifting Apart	arXiv (UC Berkeley)	Diagnoses and fixes a “double-drift” instability in diffusion policy-gradient RL via interleaved self-distillation.
vlm-pbrs-automating-potential-based-reward-shaping-vlm-guidance Automating Potential-based Reward Shaping with VLM Guidance	arXiv	Learns potential-based reward shaping from VLM image-pair preferences while preserving optimal-policy guarantees.

Humanoid Robotics

Release	Venue	Significance
humanoid-dart-diffusion-guided-augmentation-loco-manipulation Humanoid-DART: Diffusion-guided Augmentation for Humanoid Loco-Manipulation	arXiv	⭐ HIGH PRIORITY: diffusion-generated goal trajectories plus curriculum relabeling scale humanoid training data from just 4 seed demos.
deed-data-efficient-post-training-retail-humanoid-vla DEED: Data-Efficient Post-Training VLA Framework for Retail Humanoids	arXiv	⭐ HIGH PRIORITY: data-efficient post-training pipeline (RECAP-style experience refinement) deployed on a real supermarket restocking task.
skild-s1-in-context-learning-robotics Skild AI: Introducing S1 — In-Context Learning for Robotics	blog	⭐ HIGH PRIORITY: single human-video demo drives zero-fine-tuning task execution, 66% vs. 9% baseline on unseen long-horizon tasks.
figure-ai-index-crowdsourced-human-video-data-platform Figure AI: Index — Crowdsourced Human Video Data Platform	blog / press	⭐ HIGH PRIORITY: gig-economy data platform reports 16M+ videos from 108 countries feeding Helix’s training pipeline, $1B committed.
wolf-vla-whole-body-humanoid-optimal-locomotion-framework WOLF-VLA: Whole-Body Humanoid Optimal Locomotion Framework	arXiv	Large optimal-control-generated locomotion dataset trains a language-conditioned whole-body VLA for physically consistent motion.
golem-modular-humanoid-autonomy-ev-battery-disassembly GOLEM: Modular Humanoid Autonomy Towards EV Battery Disassembly	arXiv	Open-source modular ROS2 system on a Unitree H1-2 disassembles a real Hyundai Ioniq 5 battery pack across three autonomy levels.
xiaomi-second-generation-humanoid-tieda-debut Xiaomi’s Second-Generation Humanoid Robot Debuts at WRC 2026	blog / press	New 66-DoF Xiaomi humanoid trained embedded in the company’s own EV factory; embodied model drives fully autonomous action selection.

---

Generated automatically. All entries verified via web search.