Back to News & Insights
Frontier Observations

CoRL, RSS, and ICRA: the "real-world turn" in top robotics-learning venues

CoRL, RSS, and ICRA: the "real-world turn" in top robotics-learning venues

2026-09-08
Chief News Officer, TOBI Technology
Frontier Observations
Data annotation platform and quality control process

Introduction

NeurIPS and ICLR focus on models; CoRL, RSS, and ICRA focus on the robots themselves. An important shift along these conference tracks in 2025–2026 is that "simulation results plus limited real-robot validation" is being replaced by "large-scale real-world experiments"—in CoRL 2025's paper categories, Sim-to-Real is already a standalone track; the "robot teaching in the wild" concept proposed by the UMI paper at RSS 2024 evolved into a full-fledged data industry by 2026; among ICRA 2026's Best Paper nominations is work on few-shot adaptation for bimanual manipulation, while tactile-related force-control papers explicitly incorporate "force" into the action space. For industry, the value of this track lies in: what it validates is often directly deployable.

Technical Highlights

CoRL: From Policy to System. According to a summary of public papers, the embodied AI topic categories at CoRL 2025 include VLA models, world models, policies, humanoids, navigation, benchmarks and datasets, dexterous manipulation, and Sim-to-Real. Notably, "benchmarks and datasets" and "Sim-to-Real" are listed side by side — reflecting the conference's dual focus on "reproducibility" and "transfer effectiveness." CoRL has long been the main venue for robot learning moving from simulation to real hardware, and the maturity of its sim-to-real track often leads industrial deployment by 12–18 months.

RSS: From Algorithms to Data Interfaces. UMI (Universal Manipulation Interface) at RSS 2024 is one of the most cited and practically influential robotics data papers in recent years. Its core idea — wearing the end-effector on a human hand and moving the rest of the robot out of the capture loop — directly spawned subsequent exoskeleton approaches (DexUMI) and a wave of startups. RSS has consistently emphasized "real-world, reproducible, physically constrained" work, which complements pure learning conferences.

ICRA 2026: Engineering details gain acceptance at top-tier conferences. Bi-Adapt, a few-shot adaptation method for bimanual manipulation (ICRA 2026, Best Paper nomination), achieves few-shot adaptation to novel categories of 3D objects through semantic correspondence; FARM (ICRA 2026) uses a modified UMI gripper integrated with a GelSight Mini to collect human demonstrations, with a policy that jointly predicts pose, gripper width, and gripping force, outperforming baselines across three task categories: high-force, low-force, and dynamic-force. What these two works have in common is:It solves the engineering problem of "almost usable, but not quite.", rather than proposing a new paradigm.

Cross-impact from ICML 2026. Although EgoScale (ICML 2026) and DreamDojo (ICML 2026) are papers from a machine learning conference, their topics—scaling up egocentric data and world models from human video—are fundamentally robot data problems, and have become important sources of citations for the CoRL/RSS community.

Key Breakthroughs

  • •Explicitly incorporating "force" into the action space: FARM demonstrates that gripping force should not be a byproduct of the gripping command, but rather a predictable and controllable output—particularly critical in the manipulation of fragile and deformable objects.
  • •Few-shot adaptation extends to bimanual manipulation and 3D objects: Bi-Adapt advances few-shot adaptation from single-arm planar objects to dual-arm three-dimensional objects, narrowing the gap with real production-line requirements.
  • •Sim-to-Real becomes a standalone and mature track: from domain randomization to real-to-sim-to-real (real2sim2real) closed-loop methods, which have become a standard toolbox rather than an exploratory direction.
  • •Haptics and force feedback gain significantly higher visibility at top robotics conferencesTactile-related papers at ICRA/IROS have moved from workshop level in 2024 to the main conference, converging with the foundation model track.

Impact on the Industry

First,Real-robot experiment scale is becoming a threshold for publication, which directly drives up demand for "reproducible data acquisition protocols"—papers need to clearly specify camera configurations, synchronization methods, calibration procedures, and failure rates, all of which are areas that data acquisition equipment vendors can standardize.

Second,Once force/tactile enters the action space, data collection must simultaneously record force and tactile channelsand ensure its temporal alignment with the visual channel. Vision-only datasets will depreciate significantly over the next two years.

Third,As few-shot adaptation capability improves, the value of "scenario coverage breadth" relative to "single-scenario data depth" increases. Clients no longer need to collect hundreds of trajectories for each new object, but the data must cover a sufficient range of object categories and semantic correspondences.

Image Suggestions

Figure 1 (placement: after the introduction)

  • •Visual description: A schematic horizontal comparison table across three conference tracks. Three columns represent CoRL / RSS / ICRA, each labeled with its core focus (system-level real-world validation / data interfaces and physical constraints / engineering details and component-level methods), with an arrow at the bottom indicating the shared trend "from simulation-first to real-world-first."
  • •Figure caption: Figure 1 | Division of labor and shared trends across three top robotics conference tracks (the division of labor is a synthesis by this article,This is an organizational judgment).

Figure 2 (placement: after the FARM section in "Technical Highlights")

  • •Figure description: Schematic of the FARM action space. The left input is vision + touch (GelSight images) and proprioception; the middle is the diffusion policy; the right output is three parallel action quantities: pose, grip width, and grip force. "Grip force" is highlighted in an accent color and annotated "Traditional approach: a byproduct of the grip command."
  • •Figure caption: Figure 2 | Explicitly incorporating force into the action space: joint prediction of pose, gripper width, and gripping force (illustrated based on the ICRA 2026 FARM work).

Figure 3 (placement: after the second item in "Impact on Industry")

  • •Visual description: A schematic of temporal alignment in multimodal data acquisition. A horizontal timeline shows four channels: multi-view RGB vision, tactile/force, proprioception, and action/trajectory. A dashed line between the vision and tactile channels is labeled "hardware synchronization required," and a magnifying glass illustrates the consequences of misaligned timestamps (contact events and force peaks offset from each other).
  • •Figure caption: Figure 3 | Once force/tactile data enters the dataset, cross-modal temporal alignment becomes a hard requirement.

Source and Date

  • •UMI: Universal Manipulation Interface(RSS 2024);DexUMI(NVIDIA,2025–2026)。
  • •CoRL 2025 paper category list (VLA / World Model / Policy / Humanoid / Navigation / Benchmark and Dataset / Dexterous Manipulation / Sim-to-Real): open-source summary repository, updated since September 2025.
  • •Bi-Adapt: Few-Shot Bimanual Adaptation for Novel Categories of 3D Objects via Semantic Correspondence (ICRA 2026, Best Paper nomination).
  • •FARM / Tactile-Conditioned Diffusion Policy for Force-Aware Robotic Manipulation (ICRA 2026, TU Darmstadt et al.).
  • •EgoScale、DreamDojo(ICML 2026)。

Annotation Notes: The final accepted paper lists for CoRL 2026 and RSS 2026 are subject to official conference announcements; the categorizations and trends cited in this article are primarily based on CoRL 2025 and publicly available summaries (specific papers for the 2026 cycle to be verified). The division of labor across conferences and trend judgments are syntheses by this article, and areSpeculative compilation.

Contact Information · WeChat QR Code · Requirements Form

Hunan TOBI Technology Co., Ltd. (TOBI) · Embodied AI Partnerships +86 137 5502 0164

WeChat QR Code

Fill out the requirements form