Weekly Research Digest — 2026-08-09
36 new entries this week across 4 topic areas. This week’s sweep surfaced an unusually large cluster of VLA/policy post-training work — 17 of the 36 entries touch test-time adaptation, human-in-the-loop correction, preference optimization, or data synthesis for VLAs, flagged ⭐ HIGH PRIORITY below.
Vision-Language-Action (VLA) Models
| Release | Venue | Significance |
|---|---|---|
| ego2robot-scalable-robot-data-synthesis-egocentric-human-data Ego2Robot | arXiv | ⭐ HIGH PRIORITY: Largest ego-to-robot synthetic dataset yet (18.5k hrs, 15 morphologies) for VLA pretraining. |
| prigo-test-time-primitive-guidance-diffusion-flow-policies PriGo | arXiv | ⭐ HIGH PRIORITY: Test-time primitive guidance steers diffusion/flow policies without retraining. |
| retrieve-then-steer-online-success-memory-test-time-adaptation-vla Retrieve-then-Steer | arXiv | ⭐ HIGH PRIORITY: Online success-memory test-time adaptation for frozen generative VLAs. |
| wizard-robotic-policy-adaptation-weight-space-meta-learning WIZARD | arXiv | ⭐ HIGH PRIORITY: Predicts LoRA weight updates in one forward pass — gradient-free VLA adaptation. |
| set-supervised-diffusion-policy-action-chunking-corrections Set-Supervised Diffusion Policy | arXiv / RSS 2026 | ⭐ HIGH PRIORITY: Contrastive supervision from DAgger-style correction pairs, peer-reviewed at RSS. |
| vla-ad-offline-semantic-guidance-efficient-vla-distillation VLA-AD | arXiv | ⭐ HIGH PRIORITY: 44x VLA distillation via offline VLM semantic supervision; headline numbers need independent checking. |
| retrieval-vla-training-free-in-context-adaptation Retrieval-VLA | CVPR 2026 | ⭐ HIGH PRIORITY: Training-free episodic-memory in-context adaptation, peer-reviewed at CVPR. |
| dypes-vla-shared-dynamics-embodiment-specific-control-cross-embodiment DyPES-VLA | arXiv | Splits shared dynamics priors from embodiment-specific control across arm/dual-arm/humanoid. |
| cross-embodiment-transfer-behavior-aligned-representations Cross-Embodiment Transfer via Behavior-Aligned Representations | arXiv | New benchmark + representation study for transferring from action-free data. |
| cloak-zero-shot-cross-embodiment-masking-end-effector Cloak | arXiv | Simple end-effector masking trick enables zero-shot gripper/hand transfer. |
| primitivevla-reusable-motion-primitives-efficient-generalizable-manipulation PrimitiveVLA | arXiv | Primitive-based “disassemble & assemble” framing matches 100%-data baseline with 50% data on LIBERO. |
| instant-fold-in-context-imitation-learning-deformable-manipulation Instant-Fold | arXiv | Single-demo in-context learning for cloth folding, sim-to-real zero-shot claim. |
| generalist-ai-gen1-thousand-hands-cross-embodiment-end-effector Generalist AI — GEN-1 “Thousand Hands” | blog | ⭐ HIGH PRIORITY: 9,000 gripper/tool variants from 500k+ hrs real data; unverified vendor claims. |
World Models for Robotics
| Release | Venue | Significance |
|---|---|---|
| feedback-world-model-precise-guidance-diffusion-policy Feedback World Model | arXiv | ⭐ HIGH PRIORITY: Closed-loop world-model guidance corrects diffusion policies online, no retraining. |
| wayve-gaia-4-multimodal-world-models-closed-loop-simulation Wayve GAIA-4 | blog | Closed-loop multimodal (video+radar) driving world model for safety evaluation at scale. |
| deform360-massive-multiview-visuotactile-dataset-deformable-world-models Deform360 | arXiv / ECCV 2026 | Large visuotactile dataset for comparing 2D vs. 3D world-model paradigms on deformables. |
| how-should-world-models-be-evaluated-decision-making-position How Should World Models Be Evaluated? | arXiv | Position paper arguing for interventional-reasoning-centric evaluation over visual realism. |
| gauge-measurement-grounded-benchmark-physical-fidelity-simulation-video-world-models GAUGE | arXiv | Measurement-grounded physical-fidelity benchmark spanning simulators and video world models. |
| vlaflow-unified-training-framework-co-training-future-latent-alignment VLAFlow | arXiv | ⭐ HIGH PRIORITY: Controlled ablation shows future-latent-alignment + language co-training beats either alone. |
| dream-tac-unified-tactile-world-action-model-contact-rich-manipulation Dream-Tac | arXiv | Tactile world-action model, +31.7% action accuracy on contact-rich tasks over vision-only. |
| geniworld-generalizable-interactive-world-model-visual-actions GeniWorld | arXiv | URDF-rendered visual-action conditioning reduces scene overfitting, doubles as policy evaluator. |
| omega-0-latent-predictive-world-action-model-concurrent-humanoid-loco-manipulation ω₀ | arXiv | Concurrent (non-hierarchical) humanoid loco-manipulation WAM + new 40hr real-world household dataset. |
Reinforcement Learning for Robotics
| Release | Venue | Significance |
|---|---|---|
| rtcf-retrieve-in-time-correct-in-frequency-test-time-correction-vla RTCF | arXiv | ⭐ HIGH PRIORITY: Causal memory alignment corrects long-horizon VLA execution error, training-free. |
| unintervene-agentic-intervention-efficient-real-world-rl UniIntervene | arXiv | ⭐ HIGH PRIORITY: Learns when to auto-intervene, cutting human supervision cost in real-world RL. |
| mapl-multi-objective-preference-learning-robot-locomotion MAPL | arXiv | ⭐ HIGH PRIORITY: LLM-generated multi-criteria preferences replace hand-engineered locomotion rewards. |
| rarm-confidence-gated-progress-reward-modeling-rl-manipulation RARM | arXiv | Reward model from a single demonstration, confidence-gated to resist reward hacking. |
| densereward-dense-reward-learning-failure-synthesis-robotic-manipulation DenseReward | arXiv | ⭐ HIGH PRIORITY: Dense per-timestep reward model trained on 27k synthesized failure trajectories. |
| progress-reward-modeling-robotic-learning-survey Progress Reward Modeling Survey | arXiv | First unified survey of the fast-growing progress-reward-modeling literature. |
| torl-vla-tactile-guided-online-reinforcement-learning-contact-rich-manipulation TORL-VLA | arXiv | ⭐ HIGH PRIORITY: Online RL refines tactile VLA actions at deployment for shifted contact conditions. |
| d-vla-high-concurrency-distributed-asynchronous-rl-framework D-VLA | arXiv | ⭐ HIGH PRIORITY: Distributed async RL training infrastructure for billion-parameter VLAs. |
| sp3o-reinforcement-learning-segment-preferences-without-reward-modeling SP3O | arXiv | ⭐ HIGH PRIORITY: Reward-model-free, critic-free segment-preference RL; robotics + LLM domains. |
Humanoid Robotics
| Release | Venue | Significance |
|---|---|---|
| teleopit-full-embodiment-humanoid-teleoperation-system Teleopit | arXiv | Full-body VR teleoperation with hand-agnostic retargeting, 90-95% success from 96 demos. |
| thorarena-benchmarking-humanoid-physical-interaction-motion-force-demonstrations ThorArena | arXiv | First benchmark jointly scoring tracking, stability, and robustness under real external force. |
| heft-heavy-payload-full-size-humanoid-teleoperation-privileged-motion-guidance HEFT | arXiv | Teacher-student control for heavy-payload (24kg on 65kg robot) full-size humanoid teleoperation. |
| humanoidarena-benchmarking-egocentric-hierarchical-whole-body-learning HumanoidArena | arXiv | Benchmark shows hierarchical humanoid control is fragile across different low-level trackers. |
| light-loco-parkour-versatile-perceptive-whole-body-locomotion-multi-skill-distillation Light-Loco-Parkour | arXiv | End-to-end onboard-only perceptive locomotion with learned, gate-free skill transitions. |
Generated automatically. All entries verified via web search.