Each section below summarizes one paper's reported method, experiments, and limits from its full review. The studies use different tasks and metrics; their results are not directly comparable.
arXiv:2611.07881
TactileLab: Sim-to-Real Dexterous Robot Manipulation with Efficient Shear-Sensitive Tactile Simulation
TactileLab presents an end-to-end sim-to-real framework for contact-rich robot manipulation that adds dense tangential-motion information to efficient rigid-body tactile simulation. Its simulated tactile flow is computed by projecting contacted object points into the tactile image plane and measuring image-space displacement, avoiding explicit soft-body elastomer simulation. Combined with contact-depth maps, PETS-Net policy adaptation, and TacFlowNet real-to-sim translation from consecutive tactile frames, the system supports scalable reinforcement-learning training and deployment across dual-arm following and dexterous in-hand rotation.
How it works
The simulator generates three classes of observations: low-dimensional contact descriptors such as force, position, and shear; a contact-depth image representing normal penetration; and a two-channel tactile-flow field representing image-space tangential displacement. Flow is computed by identifying contact pixels, projecting the associated contacted object-surface points into the tactile image plane, transforming those points by relative rigid-body motion, and taking the projected displacement.
What the paper evaluates
The paper does not rely on a named public benchmark dataset. It evaluates two simulated task benchmarks—dual-arm 6D object following and dexterous in-hand rotation—using Allegro Hand and LEAP Hand embodiments, plus a paired real-to-sim tactile dataset collected from prescribed 6D contact poses and motions. The sensor study covers markerless GelSight Mini, DIGIT, 9DTact, GelSight Mini with markers, flat TacTip, and hemispherical TacTip. Simulation reports time-to-fail, rotation counts, and fall rate; real in-hand trials report rotations and time to termination; flow grounding reports MAE and R².
Limit to keep in view
The real-to-sim transfer assumes sticking contact and does not explicitly model slip; abrupt disturbances or contact changes may therefore produce unrecoverable states.
arXiv:2611.01934
TriPhysBench: Evaluating Physical Intelligence for Embodied Agents
TriPhysBench evaluates whether video-capable agents can coordinate physical perception, external physical-knowledge retrieval, and quantitative reasoning. It contains 410 questions from 200 videos across measurement-grounded and knowledge-grounded tasks. In a six-model evaluation, Gemini-3.6-Flash has the highest full-benchmark answer score, 0.5038. The authors also introduce TriPhys, which trains on complete tool-use trajectories with supervised fine-tuning followed by direct preference optimization. On 110 questions from videos held out by video-level partition, all three trained Qwen configurations improve over their corresponding backbones; TriPhys-27B reaches 0.5818 answer accuracy, a reported 17.7% relative gain.
How it works
The benchmark contains 410 questions from 200 videos. MPR uses 90 EgoExo4D recordings and 228 questions; pose-based measurements and tracked trajectories support physical quantities, with selected observations withheld so solvers must recover them from video. KPR uses 110 Physics-aware-videos recordings and 182 questions; observations are stated, while selected external specifications such as mass or dimensions must be retrieved and matched to the described object and conditions.
What the paper evaluates
The paper first evaluates six vision-language models on the full 410-question benchmark: GPT-5.5, Gemini-3.6-Flash, InteractiveOmni-4B, Qwen3.5-4B, Qwen3.5-9B, and Qwen3.6-27B. All receive the same instructions and a native tool loop with web search and Python execution for at most six turns. Each returns a solution and final answer, which are scored separately. ACC is higher-is-better answer correctness with partial credit; SCS is higher-is-better solution correctness; OGS is their weighted aggregate. Gemini-3.6-Flash has the highest full-benchmark ACC at 0.5038 and OGS at 0.5585. GPT-5.5 has the highest MPR ACC, 0.4316, while Gemini leads KPR ACC at 0.6004.
Limit to keep in view
The adapted models are three Qwen configurations, and the training comparison uses one video-level held-out partition of 110 questions; broader model-family coverage and repeated runs are not reported here.
arXiv:2610.41190
FlowMoDE: Coarse-to-fine Flow Matching for Detail-aware Sim-to-Real Monocular Depth Estimation
FlowMoDE is a coarse-to-fine flow-matching model for monocular depth estimation that targets the tension between discriminative models' weak fine-structure recovery and generative models' high inference cost. It extracts semantic features from a frozen DINOv2-B encoder, predicts a coarse depth velocity field with conditioned Transformer blocks, then refines it at a higher spatial resolution. Trained only on 54K synthetic Hypersim indoor frames, the model uses two-step inference and reports the best average rank across 10 depth metrics on NYUv2, KITTI, ETH3D, ScanNet, and DIODE, while also improving point-cloud reconstruction and qualitative robotic-scene perception.
How it works
Flow matching learns a time-dependent velocity field that transports samples from a simple source distribution to the data distribution. The model uses conditional flow matching with a noisy depth representation and an input RGB image. The semantic encoder produces a condition map, while the timestep is jointly embedded with that semantic condition. The architecture has three decoder stages. The Semantic Decoder processes projected DINOv2-B features.
What the paper evaluates
Training uses Hypersim, a synthetic indoor dataset with 54K frames and corresponding depth maps. Evaluation follows Marigold on five real-world datasets: NYUv2, KITTI, ETH3D, ScanNet, and DIODE. Inputs are resized with a long edge of 768, and the reported metrics are absolute relative error AR↓ and threshold accuracy δ1↑. On the depth benchmarks, FlowMoDE reports AR/δ1 values of 4.0/98.1 on NYUv2, 6.2/95.9 on KITTI, 3.9/98.8 on ETH3D, 4.3/97.6 on ScanNet, and 18.8/80.4 on DIODE. Its average rank is 2.2, compared with 3.7 for UniDepth v2, 6.1 for PPD, and 6.9 for DepthMaster.
Limit to keep in view
The paper reports occasional flying pixels caused by patch embedding in the flow-matching pipeline.
arXiv:2610.39287
OpenFlyV2: An onboard vision-language-action model for aerial perception and navigation
OpenFlyV2 is a unified 1-billion-parameter vision-language-action model for aerial robots that combines a 0.8B Qwen3.5-VL backbone with a 0.2B action expert. It supports aerial perception, visual grounding, spatial reasoning, language-conditioned navigation, and continuous waypoint generation from one shared representation. Motion-Adaptive Memory compresses visual-action history, while a Motion-Structured Prior initializes flow-based action generation from recurring trajectory patterns. Trained only on simulated data from OpenFlyV2-130K, the model generates 12-waypoint action chunks at 10 Hz on an NVIDIA Jetson Orin NX and transfers zero-shot to physical flight, reaching 48% success across 100 trials in four real-world environments.
How it works
The architecture combines a 0.8B Qwen3.5-VL backbone with a 0.2B action expert. The backbone receives the current aerial image, instruction, robot state, and MAM tokens. It can either decode a language answer or condition waypoint generation. The action representation is a 12-step chunk with three body-frame motion increments and a stop signal. MAM forms a bounded history from causal motion events.
What the paper evaluates
The evaluation covers aerial vision-language understanding, simulated VLN, ablations, onboard efficiency, and physical flight. Vision-language evaluation used 300 examples each for color recognition, object description, distance estimation, and spatial reasoning, plus 1,000 visual-grounding examples. Simulated navigation used 800 seen and 300 unseen episodes, with success requiring an explicit stop within 25 m of the goal. The baselines were AeroVLA, DualVLN, OpenUAV, and OmniNav.
Limit to keep in view
Physical evaluation covered 100 manually judged trials, and the aggregate success rate was 48%; the field criterion also differs from the 25-meter metric-distance success rule used in simulation.
Read the full reviews
- TactileLab: Sim-to-Real Dexterous Robot Manipulation with Efficient Shear-Sensitive Tactile Simulation · arXiv:2611.07881
- TriPhysBench: Evaluating Physical Intelligence for Embodied Agents · arXiv:2611.01934
- FlowMoDE: Coarse-to-fine Flow Matching for Detail-aware Sim-to-Real Monocular Depth Estimation · arXiv:2610.41190
- OpenFlyV2: An onboard vision-language-action model for aerial perception and navigation · arXiv:2610.39287
Source note: findings and limitations above come from the linked full paper reviews. Follow each original arXiv paper for the authors' complete methods and results.