Research Landscape
Today's batch reveals a strong focus on scaling robot learning through better data collection and transfer mechanisms. Papers like RoSHI, TAMEn, and BiDexGrasp attack the data bottleneck from complementary angles—wearable sensing, tactile-aware systems, and large-scale grasp annotation—suggesting the field is moving beyond algorithmic breakthroughs toward systematic solutions for data scarcity. These efforts directly enable downstream work in policy transfer (Learning-Based Assembly, Sustainable Transfer) and sim-to-real robustness (Robust Quadruped Locomotion via evolutionary RL).
Multi-robot coordination continues to mature with dual approaches gaining traction: classical declarative and rule-based methods (Aggregate Programming, Logical Robots) coexist with learning-based planners (Train-Small Deploy-Large, Differentiable Environment-Trajectory Co-Optimization). This reflects pragmatism—formal methods provide safety guarantees for swarms, while diffusion models and bi-level optimization unlock flexibility in dynamic, partially observable environments. The fact that both lineages advance simultaneously suggests neither dominates the design space.
Vision-language models have graduated from pure task execution toward autonomous self-diagnosis: KITE demonstrates that tokenized, keyframe-anchored evidence can substantially improve VLM-based failure detection beyond vanilla Qwen2.5-VL. Simultaneously, scene understanding pipelines (Genie Sim PanoRecon, MoRight) enable better 3D reconstruction and disentangled motion control, bridging perception and planning. Infrastructure requirements (RTK-SLAM Dataset, CADENCE energy-aware sensing) and foundational architectures (AEROS, RichMap) round out the ecosystem, signaling that 2026 robotics is as much about systems integration and measurement rigor as novel learning algorithms.
Vision-Language Models & Scene Generation
VLM-based failure analysis, motion control, 3D reconstruction
3
Multi-Robot Coordination & Planning
Aggregate programming, diffusion planners, declarative multi-agent
4
Manipulation & Grasping
Bimanual grasping, tactile sensing, peg-in-hole assembly
5
Motion Planning & Robot Architecture
Flow matching, OS design, quadruped locomotion
4
SLAM, Localization & Autonomous Driving
Visual SLAM, RTK positioning, trajectory prediction
4
Human Data, Bio-Inspired & Infrastructure
Wearable sensing, proprioceptive joints, telecom world models
5
Vision-Language Models & Scene Generation
Core Contributions
- Solves long-context video bottleneck by tokenizing only motion-salient keyframes with BEV representations, enabling VLMs to diagnose failures without expensive training
- Substantially outperforms vanilla Qwen2.5-VL on RoboFAC, with particularly strong gains in simulation failure detection and localization tasks
- Training-free front-end that transforms robot execution videos into compact evidence—first principled approach to keyframe-anchored VLM prompting for robotics failure analysis
Show Abstract
▼
Training-free, keyframe-anchored front-end that converts long robot-execution videos into compact tokenized evidence for VLMs. Uses motion-salient keyframes with BEV representations. On RoboFAC benchmark, substantially improves over vanilla Qwen2.5-VL especially in simulation failure detection/identification/localization.
Core Contributions
- First framework to disentangle object motion from camera viewpoint, decomposing control into active (user-driven) and passive (consequence) components
- Supports both forward control (user commands) and inverse reasoning (predict user intent from observed motion), enabling more natural human-robot interaction
- State-of-the-art on three benchmarks with a unified architecture—demonstrates that explicit factorization of motion intent improves generalization over end-to-end approaches
Show Abstract
▼
Framework for disentangled motion control: separate object motion and camera viewpoint. Decomposes motion into active (user-driven) and passive (consequence) components. Supports forward and inverse reasoning. State-of-the-art on three benchmarks.
Core Contributions
- Feed-forward Gaussian-splatting pipeline overcomes cold-start problem in simulation—reconstructs full 3D scenes from single panoramic images in seconds
- Depth-aware fusion strategy integrates multiple sensor modalities, critical for manipulation task realism in simulated training
- Direct integration into Genie Sim reduces sim-to-real friction for manipulation by enabling rapid scenario prototyping without manual asset creation
Show Abstract
▼
Feed-forward Gaussian-splatting pipeline for 3D scene reconstruction from panorama. Depth-aware fusion strategy. Integrated into Genie Sim for manipulation tasks.
Multi-Robot Coordination & Planning
Core Contributions
- First production-grade application of aggregate programming to multi-robot coordination—demonstrates scalability and fault tolerance in realistic university library environment
- Combines field calculus abstractions with validated simulation and hardware experiments, proving formal approaches work beyond toy domains
- Provides templates for adaptive swarm behaviors that handle decentralized coordination without centralized planning—key advantage for large fleets
Show Abstract
▼
Multi-robot systems using Aggregate Programming for coordination in a university library prototype, validated with simulations and real tests.
Core Contributions
- Bi-level optimization framework co-designs safe trajectories and environment layouts jointly—first approach to make environment configuration differentiable via KKT + Implicit Function Theorem
- Novel safety metric grounded in measure theory enables principled quantification of collision risk across multi-agent systems
- Enables discovery of non-intuitive, safer environment configurations automatically—useful for infrastructure design and autonomous systems deployment
Show Abstract
▼
Bi-level optimization: lower-level trajectory optimization + upper-level environment optimization. Uses KKT + Implicit Function Theorem. Novel safety metric via measure theory.
Core Contributions
- Leverages Logica (Google's declarative query language) for multi-agent simulation—maps logic predicates directly to motor outputs, eliminating imperative control code
- Enables humans and AI to specify swarm behaviors as logical constraints rather than sequential scripts, reducing specification errors
- Demonstrates alternative to code-based multi-agent programming, opening robotics to domain experts unfamiliar with traditional programming
Show Abstract
▼
Multi-agent simulation with declarative behavior in Logica. Logic predicates map observations to motor outputs.
Core Contributions
- Diffusion model planners generalize across swarm sizes without retraining—train on 2-3 agents, deploy on 5-10, addressing scalability bottleneck
- Inter-agent attention + temporal convolution architecture captures both spatial interactions and temporal dynamics elegantly
- Enables rapid deployment to larger teams without computational cost of retraining, critical for field robotics applications
Show Abstract
▼
Diffusion model planner for varying agent numbers. Trained on few agents, generalizes to more. Inter-agent attention + temporal convolution.
Manipulation & Grasping
Core Contributions
- MPC-based shared teleoperation framework with virtual object method simplifies multi-object constraint handling—operator controls aggregate motion, not individual contacts
- 72.45% reduction in sliding distance and complete elimination of tip-overs (0% vs 13.9% baseline) through force-aware control
- Demonstrates practical path to non-prehensile multi-object tasks, relevant for warehouse automation and unstructured environments
Show Abstract
▼
MPC-based shared teleoperation for multi-object nonprehensile transport. Virtual object method for constraint simplification. Reduces sliding distance by 72.45%, eliminates tip-overs (0% vs 13.9%).
Core Contributions
- Cross-morphology wearable interface enables cost-effective, robot-agnostic tactile data collection—solves sensor cost bottleneck that limits grasp dataset scale
- Dual-modal pipeline (precision + portable) with pyramid data regime increases task success from 34% to 75%, demonstrating tactile feedback is learnable and valuable
- First large-scale closed-loop tactile data collection system—addresses why contact-rich tasks remain hard despite vision-based datasets
Show Abstract
▼
Cross-morphology wearable interface for tactile data collection. Dual-modal pipeline (precision + portable). Pyramid data regime. Increases task success from 34% to 75%.
Core Contributions
- Residual RL with composite skills (pre/post/invariant conditions) enables task adaptation without monolithic retraining—modular approach to assembly robustness
- Demonstrates SAC+JAX integration on real UR5e peg-in-hole, bridging sim-to-real with structured skill composition
- Composite skill framework provides interpretability—domain experts can reason about which conditions must hold for successful assembly
Show Abstract
▼
Residual RL for peg-in-hole assembly with composite skills. Pre/post/invariant conditions. Evaluated on UR5e with SAC+JAX.
Core Contributions
- Demonstrates policy transfer across heterogeneous robot platforms for peg-in-hole, addressing generalization concerns in embodied learning
- Fine-tuning significantly outperforms zero-shot transfer, quantifying the benefit-cost tradeoff of domain adaptation
- Enables skill libraries to be shared across platforms, reducing training overhead when deploying to new hardware
Show Abstract
▼
Policy transfer across robot platforms for peg-in-hole. Fine-tuning significantly improves over zero-shot transfer.
Core Contributions
- Large-scale bimanual grasp dataset (6351 objects, 9.7M annotations) fills a critical gap—most prior work focuses on single-arm, limiting applicability to dual-arm systems
- Two-stage synthesis (region-based initialization + force-closure optimization) provides computational efficiency and physical validity
- Bimanual coordination module enables grasp quality assessment across morphologically distinct hand pairs, useful for heterogeneous manipulation teams
Show Abstract
▼
Large-scale bimanual grasp dataset: 6351 objects, 9.7M grasp annotations. Two-stage synthesis: region-based init + force-closure optimization. Bimanual coordination module.
Motion Planning & Robot Architecture
Core Contributions
- Open-loop end-to-end neural planner using flow matching generates multi-modal trajectories in one forward pass—avoids iterative sampling bottleneck
- Best-of-N sampling provides flexible accuracy/speed tradeoff—operator can increase N for tighter paths during final approach
- Demonstrates flow matching (less understood than diffusion) is viable for continuous control, expanding toolbox for generative robot planning
Show Abstract
▼
Open-loop end-to-end neural motion planner using flow matching for multi-modal path generation. Best-of-N sampling improves planning success and efficiency.
Core Contributions
- High-precision reachability map achieves >98% accuracy with only 1-2% false positives and ~15μs query latency—enables real-time planning constraints
- MMD metrics quantify workspace similarity across embodiments, enabling direct reachability map transfer with 26% improvement in diffusion policy performance
- Solves a practical deployment problem: how to reuse inverse kinematics knowledge across robot variants without recomputing
Show Abstract
▼
High-precision reachability map: >98% accuracy, 1-2% false positives, ~15μs/query. MMD metrics for workspace similarity. Up to 26% improvement in cross-embodiment diffusion policy transfer.
Core Contributions
- Runtime OS for robots with pluggable Embodied Capability Modules enables modular deployment—100% task success vs 67-93% for integrated baselines
- Zero false acceptances in policy enforcement demonstrates robust containment—modules cannot silently violate safety constraints
- Provides missing OS-level abstraction for embodied AI, analogous to how Linux changed general computing—enables rapid capability composition
Show Abstract
▼
Runtime OS for robots with installable Embodied Capability Modules. 100% task success vs 67-93% baselines. Zero false acceptances in policy enforcement. Franka Panda evaluation.
Core Contributions
- CEM-TD3 hybrid achieves 19574.33 mean reward on rough terrain vs -99.73 for vanilla TD3—evolutionary strategy discovers better initialization and exploration
- Evolutionary variants retain capability under terrain transfer, demonstrating evolutionary search finds more robust policies than pure gradient descent
- Addresses long-standing challenge in quadruped learning: why gradient-based RL struggles on unstructured terrain despite apparent convexity
Show Abstract
▼
CEM-TD3 achieves 19574.33 mean reward on rough terrain vs -99.73 for TD3. Evolutionary variants retain capability under terrain transfer.
SLAM, Localization & Autonomous Driving
Core Contributions
- VGGT front-end with geometry-grounded transformer improves feature matching robustness, addressing the long-standing limitation that ORB-SLAM scales poorly in low-texture environments
- DEM-based graph backend + DINOv2 embeddings achieve state-of-the-art SLAM accuracy by integrating semantic and geometric constraints
- Restores high-cadence local bundle adjustment, critical for real-time applications where drift accumulates quickly
Show Abstract
▼
VGGT front-end + DEM-based graph + DINOv2 embeddings. Restores high-cadence local bundle adjustment. State-of-the-art SLAM accuracy.
Core Contributions
- Geodetic total station ground truth (not GNSS) enables centimeter-level accuracy validation where GPS fails—solves evaluation gap for urban/indoor robots
- Reveals that SE(3) alignment underestimates error by up to 76%, demonstrating common evaluation protocol is fundamentally flawed
- Dataset enables honest benchmarking of multi-sensor fusion systems in realistic degraded scenarios, critical for autonomous vehicles
Show Abstract
▼
Dataset with geodetic total station ground truth (not GNSS). SE(3) alignment underestimates error by up to 76%. Centimeter-level accuracy outdoors, decimeter indoors.
Core Contributions
- Adaptive depth estimation scales computational cost based on navigation context—75% energy reduction on edge hardware (Jetson Orin Nano) without sacrificing accuracy
- 7.43% navigation accuracy improvement demonstrates that selective refinement is beneficial, not just cost-saving
- Enables deployment to resource-constrained platforms, critical for swarms and long-endurance missions
Show Abstract
▼
Adaptive system scaling depth estimation complexity based on navigation needs. 75% energy reduction, 7.43% navigation accuracy improvement on NVIDIA Jetson Orin Nano.
Core Contributions
- Pure Transformer (no RNNs) with two-track architecture jointly predicts trajectories and behavioral intentions, eliminating decoupling errors
- Residual offset learning discovers trajectory groups self-supervised, reducing annotation burden for motion datasets
- Applies to autonomous driving prediction, enabling better anticipation of multi-modal vehicle futures without explicit mode labels
Show Abstract
▼
Pure Transformer for trajectory + intention prediction with two-track design. Learns ordered trajectory groups via residual offsets.
Human Data, Bio-Inspired & Infrastructure
Core Contributions
- Telecom World Model architecture applies learned, action-conditioned, uncertainty-aware dynamics modeling to 6G network slicing—bridges embodied AI and telecom systems
- Three-layer architecture unifies digital twins, foundation models, and planning—demonstrates that world models generalize beyond robotics
- Proof-of-concept on network slicing shows practical value for infrastructure optimization, opening robotics methodologies to telecom domain
Show Abstract
▼
Telecom World Model architecture for learned, action-conditioned, uncertainty-aware modeling of telecom dynamics. Three-layer architecture for 6G. Proof-of-concept on network slicing.
Core Contributions
- Hybrid wearable (IMUs + Project Aria glasses) estimates full 3D pose and body shape from egocentric view, solving the cold-start problem for humanoid policy learning
- Outperforms previous egocentric baselines and matches SAM3D, demonstrating that sensor fusion beats single-modality approaches for in-the-wild capture
- Enables cost-effective on-the-job human motion capture for robotics—reduces instrumentation burden for data collection in real environments
Show Abstract
▼
Hybrid wearable fusing IMUs with Project Aria glasses for full 3D pose/body shape estimation. Outperforms egocentric baselines, comparable to SAM3D. Demonstrated for humanoid policy learning.
Core Contributions
- Biomimetic joint with Type I receptor analog achieves <2 degree average error in 3D bending and twisting—validates decades of neuroscience theory in hardware
- Suggests joint receptors play greater proprioceptive role than previously thought, shifting understanding of sensorimotor control architecture
- Opens path to biologically-inspired sensing in robots, potentially simpler and more robust than vision-based proprioception
Show Abstract
▼
Biomimetic joint mimicking Type I joint receptors achieves <2 degrees average error in bending and twisting. Suggests joint receptors play greater role in proprioception than previously thought.
Core Contributions
- Argues for infrastructure-first approach to embodied AI deployment in resource-limited settings—prioritizes grid power, compute, connectivity over algorithms
- Outlines practical requirements for scaling embodied intelligence beyond well-resourced labs, addressing a critical gap in robotics deployment literature
- Emphasizes that robotics accessibility requires infrastructure investment, not just algorithmic innovation—reshapes how we should think about global impact
Show Abstract
▼
Argues for infrastructure-first approach to Embodied AI for Science. Outlines requirements for deploying embodied intelligence at scale in resource-limited settings.
Core Contributions
- RL-based RoboFish autonomously evaluates fish behavior models through closed-loop interaction—novel approach to model validation that doesn't require labeled data
- Neural network fish model shows smallest sim-to-real gap versus other learned and hand-crafted models, suggesting neural approaches capture ethology better
- Demonstrates robots can serve as experimental platforms for behavioral science, inverting typical application direction
Show Abstract
▼
RL-based RoboFish evaluates fish behavior models through closed-loop interaction. Neural network fish model shows smallest sim-to-real gap.