Embodied AI 101

Finetunes a single Vision-Language-Action (VLA) foundation model so that any robot in a team can control any other. Outperforms both per-robot specialists and a monolithic centralized policy while scaling to large teams.

Trains a compositional world model on real robot data to enable closed-loop policy improvement via future prediction and progress evaluation, bypassing both risky real-world RL and traditional sim-to-real gaps.

Open-source 3B unified embodied foundation model trained on 1.5M interleaved vision-text-action samples for perception, planning, and acting.

A single vision-language model that unifies reasoning, task planning, and human-robot interaction for complex instructions and long-horizon tasks.

Presents a neural network-based world model for model-based reinforcement learning in robotics, focusing on sim-to-real transfer for quadrupedal and humanoid robots. Enables robust policy optimization through learned environment simulation.

Embodied egocentric simulation framework that controls first-person worlds with 3D human motion and customizes evolving scenes via pose-anchored views.

Few-step auto-regressive diffusion model that converts broken 3D reconstructions into fully realized scenes, outperforming prior methods by 1-3 dB PSNR.

Examines memorization phenomena that occur during deployment of foundation model agents in practical applications.

Focuses on security aspects of deep neural networks, providing taxonomy and mitigation strategies for adversarial attacks.

First comprehensive framework for evaluating LLM mediators in real-time, emotional, socio-cognitive scenarios.

Improves embedding quality without extra training by using the unembedding matrix as a feature lens for text embeddings.

Presents an autonomous system for formalizing mathematical proofs, specifically targeting Erdős problems. Demonstrates automated proof formalization capabilities in the Lean theorem prover.

Provides a comprehensive taxonomy, benchmarks, and future directions for autonomous AI research agents. Examines systematic approaches to developing agents capable of conducting independent research.

Omnimodal world models explicitly designed for Physical AI and robotics applications. Enables improved simulation and control for robotic systems through multimodal understanding.

GPT-style Transformer trained on 2 billion motion frames enabling zero-shot dynamic humanoid control for tasks like soccer, dancing, and digging on real Unitree G1 robots without fine-tuning. Breaks the agility-generalization trade-off in humanoid robotics.

New IROS benchmark providing 500+ teleoperation episodes and physically accurate simulation assets for training policies that outperform human origami experts.

First foundation model for zero-shot grasping trained on billions of simulated grasps, enabling generalized manipulation without task-specific training.

Benchmark study examining when deep reinforcement learning outperforms calibrated baseline methods in adaptive resource control tasks.

Explores training dynamics of deep neural networks through a statistical lens, examining the duality between optimization and inference perspectives.

Vision-language model that adds generative depth prediction during pre-training for physical grounding; achieves SOTA on embodied benchiments and transfers directly to real-robot tasks.

Presents a 3D point-cloud-based world model trained on mixed real/sim data that enables zero-shot grasping and articulated object handling on real robots by explicitly modeling spatial structure.

NVIDIA Research presents LocateAnything, a unified generative grounding and detection framework based on Parallel Box Decoding (PBD). Unlike prior VLMs that serialize bounding boxes into sequential coordinate tokens, PBD treats each box as an atomic unit and predicts all coordinates in a single forward pass. This preserves intra-box geometric coherence while achieving 2.5x faster decoding throughp...

Replaces quadratic softmax attention in looped architectures with linear/sparse mechanisms for iterative memory refinement, achieving parity with standard looped transformers at much lower cost.

Shows that modern transformers are highly heterogeneous across layers and proposes layerwise learning rates based on weight spectrum shape, yielding up to 1.5× training speedup on LLaMA/GPT-style models.

Trains generalist policies in simulation on procedurally generated tools to move objects, enabling real-world tool use across varied shapes/sizes.