It’s the demos: A Deep Look at the Role of Demonstration Quality in Imitation-Based Robot Manipulation @ RSS 2026

Imitation-based robot manipulation has entered a phase where short to mid horizon tasks that can be unambiguously teleoperated have a high likelihood of success. Further Progress is increasingly driven by large-scale demonstration collection in real-world settings rather than purely algorithmic novelty. Modern pipelines, ranging from behavior cloning to diffusion- and transformer-based policies, often rely on thousands to millions of frames of demonstrations gathered via teleoperation from expert users. Within this regime, demonstration quality is a first-class variable: differences in operator skill and style (“robot wizards”), interface ergonomics, latency, action representations, and curation choices can decisively shape downstream policy performance, robustness, and generalization. Despite this, the community lacks a shared vocabulary, metrics, and engineering playbook for what “high-quality” demonstrations mean across tasks and embodiments, and how to reliably produce them at scale. The workshop targets researchers and practitioners who build, use, or maintain demonstration datasets for manipulation learning. The intended audience includes: (i) method developers working on learning-from-demonstration for dexterous manipulation, mobile manipulation, and bimanual tasks; (ii) dataset builders and infrastructure engineers responsible for logging, synchronization, calibration, and labeling; and (iii) practitioners deploying imitation-based manipulation in industrial or field environments where reliability, repeatability are central. We expect strong relevance for groups collecting demonstrations “in the wild,” where demonstrations are produced by multiple operators with varying expertise, under time pressure, and with imperfect sensing, occlusions, or changing scene conditions. Presenters and panelists will be drawn from multiple sub-communities that rarely meet in a single forum: robot manipulation learning (BC, diffusion/transformer policies, hybrid IL/RL), human-in-the-loop systems (teleoperation, XR interfaces, shared autonomy and interventions), dataset and benchmark design (metadata standards, evaluation protocols, reproducibility), and industry robotics (quality assurance, safety practices, scalable operations, and documentation). The workshop will explicitly connect “front-end” data collection decisions to “back-end” learning and evaluation outcomes, encouraging speakers to share concrete lessons from building and stress-testing real pipelines. The workshop’s relevance and impact is to complement RSS main-track research by focusing on best practices and practical failure modes that are often omitted from papers but strongly determine success in practice. Key themes include: Best-practices for real world data collection, how to train operators; how to characterize and manage operator-to-operator variation; how to design collection protocols that reduce reliance on rare “wizard” skill; how to define and measure quality beyond task success (e.g., coverage, consistency, uncertainty, recoveries, and intent clarity); how to implement quality-aware filtering, weighting, and dataset documentation; and how to report dataset properties in ways that make results comparable across labs. The expected outcome is a community-curated set of actionable recommendations, protocols, checklists, as well as open problems. The final goal is to improve reproducibility, reduce wasted data collection, and accelerate reliable imitation-based manipulation progress in 2026 and beyond.

Call for Papers

We invite participants to submit short contributions (3 pages extended abstract, unlimited references and appendix) focused on practical lessons from demonstration collection and curation: what failed, what worked, what metrics or checks were useful, and what factors mattered.

Submissions will be accepted primarily for poster presentation, with a curated subset invited for lightning talks (7 min + 3 min Q&A) emphasizing actionable takeaways and an “anecdote → hypothesis” framing. We will reserve a substantial fraction of lightning slots for students, postdocs, and first-time RSS participants.

Contributions are encouraged, but not required, to be original. The review process will be single-blind (submitted papers do not need to be anonymized). Accepted abstracts will be made available on the workshop website but will not appear in the official RSS proceedings.

Important Dates

Submission portal and exact dates will be announced here. Check back soon.