§ Robotics

LG Electronics, LG Innotek, and Hyundai Motor Group advance Nvidia robotics partnerships with data factory ops and autonomous driving infrastructure.

Honda targets Level 4 autonomy by 2027, plans AI-based map replacement by 2030 for Asian road navigation.

HHTD 2026 showcases robotics + AI infrastructure focus; Foxconn, KUKA, ABB, Yaskawa, TDK, Hiwin, TM Robot, TECO participate.

Nvidia projects 69% market share in 2026 AI robotics chips, outpacing Intel, Qualcomm, AMD via software ecosystem.

Gym-Anything for Robotics uses coding agents to automatically generate realistic physics simulations for embodied AI training at scale.

BURGER benchmark evaluates vision-language-action model generalization via scalable prompt-driven robotic drawing in simulation and real robots.

Value-of-information framework for efficient memory retrieval and credit assignment in RL, with learned world models for latency-constrained robotics.

CoC-CFG: classifier-free guidance method to calibrate reasoning-trace influence in VLA policies via learned step-level action-value weighting.

LiDAR-SHIELD provides a Dockerized adversarial robustness testing framework for 3D LiDAR perception models in autonomous driving and robotics.

AegisFlow enables training-free path-safe guidance for diffusion and flow matching models via non-myopic constraint satisfaction in image editing and robotics.

Kairos orchestrates heterogeneous PDDL planning engines for real-time robotics, selecting engines by task semantics to cover STRIPS through PDDL+ while maintaining latency.

Movement-primitive action chunking for imitation learning reduces jerkiness in robot policies by explicitly modeling temporal structure across diffusion and vision-language backbones.

UMI-Live is an open-source bimanual manipulation interface with integrated sensing and onboard compute for in-the-wild robot learning data collection without external calibration.

WorldFlow jointly predicts future video frames and 3D object trajectories via mixture-of-tokens for embodied AI planning.

Online RL framework trains lightweight UI navigation agents directly on live iOS apps via planning and grounding components.

CuraEval systematically benchmarks demonstration selection strategies for robot imitation learning on mixed-skill robomimic data.

Vision-Language-Action models trained on 1616 underwater manipulation episodes for domain-shifted robotic generalization.

Study isolates five joint-space interfaces and their filtering/scaling properties in robot learning, revealing how action-space labels mask implementation details affecting attribution.

ThinkWorld framework grounds VLA robotic learning in world dynamics and RL imagination, scaling beyond action-labeled demos.

Constraint-Referenced Strategy metric evaluates robot manipulation by separating progress along object joints from wasted motion.

Adaptive Residual Correction uses sparse human feedback to guide RL refinement of imitation policies on localized failure regions.

Vision-language models struggle to diagnose bimanual tool failures without access to low-level sensing channels beyond vision.

RL controller for fixed-wing drone interception of high-altitude targets; military application with limited commercial relevance.

300+ Boston Dynamics employees oppose sale of Spot robot dogs to ICE amid ethical concerns over law enforcement deployment.

FLORE framework learns force-reactive humanoid manipulation from robot-free demos using diffusion policies with vision, proprioception, and contact wrenches.

EBench evaluates GPT-6-Astra as a frontier MLLM agent controlling robots via direct control vs. policy harness, benchmarking capabilities against post-trained VLA and world action models.

Zero-real Agent Policy (ZAP) bootstraps a 27B vision-language model with pretraining for task understanding, then sim-trains on robot control via tool harness, achieving 96.7% real-world success zero-shot.

Routed ensemble approach for multi-objective RL enables sample-efficient per-preference policy learning across alignment and robotics tasks.

TrySee: modular pipeline integrates VLM reasoning with domain-specific perception modules for closed-loop robot task planning in specialized domains.

Tesla shifts focus from EV sales to autonomy subscriptions as revenue driver; not AI research or deployment.

Bi-Mamba architecture predicts human trajectories from noisy RGB-D sensor data for mobile robots, addressing train-deploy distribution gap.

Nvidia CEO demonstrates Microduck toy robots with voice-command capability; 20,000 units sold.

SceneBloom combines diffusion-based layout generation with geometric compilation for realistic, simulation-ready 3D indoor environments for embodied agent training.

DexPort portable interface enables dexterous hand demonstration collection using same hardware for collection and robot deployment, reducing retargeting overhead.

Backtrack: lightweight failure detection and recovery for frozen VLA models without retraining, using only a small prediction head on frozen representations.

Federated VLA adaptation framework addresses safety risks when sharing human corrections across robots, tasks, and environments without centralizing data.

Regularised RL algorithms improve sim-to-real robotic transfer vs. PPO; soft MellowMax Actor-Critic analysis shows practical robustness benefits.

VERID infrastructure couples AI reasoning with reliable laboratory automation and self-play iteration for autonomous biological discovery.

Proposes Sensorimotor Alignment as offline proxy for evaluating robotic foundation models by testing convergence between vision, language, and motion representations.

RaSFNet: ray-aligned fusion network for LiDAR semantic segmentation in autonomous driving and robotics perception.

WholeBodyMem: perceptual memory module for humanoid loco-manipulation via object tracking and cross-attention action expert.

Study compares semantic retrieval and LLM-based strategies for tool routing in robotic systems, addressing capability selection for compound natural-language commands.

PAPAV framework unifies multimodal embodied agents across computer-use and robot-use via five capability functions: perceive, plan, act, process feedback, evolve.

AtomicRL post-trains VLA models for long-horizon manipulation via atomic task structures, solving credit-assignment and subtask-imbalance bottlenecks in RL.

Evaluates frontier AI models (Claude, GPT-4, others) on LIBERO-10 robot manipulation via closed-loop visual control without task-specific training.

SP-HLM couples sparse-pose interface with learned motion priors and vision-language-action models for humanoid loco-manipulation control.

FaultGene uses LLM-assisted synthesis to generate fault-augmented test scenarios for hardware-in-the-loop validation of autonomous driving cyber-physical systems.

AdaDex combines reinforcement learning with motion imitation to train dexterous multi-finger robot hands on generalizable, human-like manipulation skills.

UniWALM unified world-action-language model learns video, action, and language jointly for embodied robot control with memory-dependent behavior.

IndustrialRemembr extends NVIDIA ReMEmbR spatiotemporal memory system to robot fleets with on-demand scene encoding to reduce compute costs.