Skip to content

Literature

CONCERTO and CHAMBER build on three tiers of external evidence: (a) peer-reviewed publications, (b) standards and technical specifications, and (c) industry signal and horizon-scanning material. The three tiers are kept separate because they have different evidentiary weight and different update cadences. Tier (a) — the peer-reviewed bibliography below — anchors the method's theoretical claims. Tier (b) lives in standards.md: binding standards (ISO, IEC, IEEE, 3GPP) that bound what "safe" and "deterministic" mean in deployment. Tier (c) lives in adr/international_axis_evidence.md: deployment signal (humanoid factory pilots, logistics fleets, surgical robotics) used as a prior on which heterogeneity axes are commercially load-bearing, but never as a substitute for the empirical ≥20pp gap rule (see ADR-007 §Validation criteria). This page covers tier (a) only.

Every entry below carries a BibTeX cite-key in square brackets matching the canonical docs/reference/refs.bib file. Cite by key from public docs and ADRs; do not introduce a citation without adding it to refs.bib first. See CONTRIBUTING.md § Bibliography hygiene for the maintainer checklist.

Taxonomy

mindmap
  root((Literature))
    AHT and ZSC
      Stone 2010
      Mirsky 2022
      Other-Play 2020
      MCS 2024
      Liu 2024 RSS
      COHERENT 2024
    Safe control / CBF
      Ames 2017
      Ames 2019
      Wang 2017
      OSCBF 2025
      Survey 2024
    Conformal prediction
      Shafer Vovk 2008
      Angelopoulos 2023
      Huriot Sibai 2025
    Benchmarks and data
      ManiSkill v3
      COLOSSEUM 2024
      DROID 2024
    Reproducibility
      Henderson 2018
      Agarwal 2021
      Jordan 2020

The five clusters below each open with a two-paragraph note locating the cited work relative to CONCERTO's claims, followed by the bibliography itself: one bullet per work, prefixed with its code-spanned [bibkey] from docs/reference/refs.bib. Where a citation also appears in an existing ADR, the ADR reference is given inline.

1. Ad hoc teamwork and zero-shot coordination

Ad hoc teamwork (AHT) and zero-shot coordination (ZSC) define the problem CONCERTO is built to solve: an ego agent must cooperate with partners whose policies, training history, and internal state are opaque at deployment time. Stone et al. (2010) [stone2010adhoc] coined the AHT formalism; Mirsky et al. (2022) [mirsky2022survey] survey a decade of subsequent work and expose the under-explored manipulation tier that CHAMBER is built to fill. Hu et al. (2020) [hu2020otherplay] introduce other-play as the first principled ZSC method, and Rahman et al. (2024) [rahman2024mcs] sharpen partner-set construction with Minimum Coverage Sets — the formal counterpart to the entropy-filtered partner zoo specified in ADR-009.

Liu et al. (2024) [liu2024llm_aht] and COHERENT (Liu et al. 2024) [coherent2024] are the two closest precedents to CONCERTO on the heterogeneous, black-box partner axes. Both stop short of contact-rich manipulation with a formal safety bound; the CONCERTO-vs-precedents table in the project README makes the differentiation explicit. They are listed here for completeness so this page can stand alone as the literature reference.

  • [stone2010adhoc] Stone, Kaminka, Kraus, Rosenschein (2010), "Ad Hoc Autonomous Agent Teams: Collaboration without Pre-Coordination," AAAI. paper
  • [mirsky2022survey] Mirsky, Carlucho, Rahman, Fosong, Macke, Sridharan, Stone, Albrecht (2022), "A Survey of Ad Hoc Teamwork Research," EUMAS. paper
  • [hu2020otherplay] Hu, Lerer, Peysakhovich, Foerster (2020), "'Other-Play' for Zero-Shot Coordination," ICML. paper
  • [rahman2024mcs] Rahman, Cui, Stone (2024), "Minimum Coverage Sets for Training Robust Ad Hoc Teamwork Agents," AAAI. paper
  • [liu2024llm_aht] Liu, Stella, Stone (2024), "LLM-Powered Hierarchical Language Agent for Real-time Human-AI Coordination," RSS. Cited throughout the CONCERTO ADRs as "Liu 2024 RSS".
  • [coherent2024] Liu, Tang, Wang, Wang, Zhao, Li (2024), "COHERENT: Collaboration of Heterogeneous Multi-Robot Systems with Large Language Models," arXiv:2409.15146. paper

2. Safe control and control barrier functions

Control barrier functions (CBFs) underwrite CONCERTO's safety claim. Ames et al. (2017) [ames2017cbfqp] is the canonical CBF-QP paper and the right starting point for any reader new to the formalism; Ames et al. (2019) [ames2019cbfsurvey] is the broader theory-and-applications survey that catalogues exponential, high-relative-degree, and discrete-time extensions used throughout ADR-004. Wang, Ames, and Egerstedt (2017) [wangames2017] extend the CBF-QP to multi-robot collision avoidance and are the direct ancestor of the per-pair CBF budget split implemented in concerto.safety.budget_split.

Morton and Pavone (2025) [morton2025oscbf] introduce the operator-splitting CBF (OSCBF), which serves as CONCERTO's inner per-arm filter, and Guerrier et al. (2024) [guerrier2024lcbfsurvey] survey the rapidly growing learning-to-CBF literature — useful when reading CONCERTO against the wider field of safety-augmented reinforcement learning. The cross-cutting multi-agent safety survey carried by ADR-004 / ADR-014 tier-2 note 45 is [lindemann2024safety].

  • [ames2017cbfqp] Ames, Xu, Grizzle, Tabuada (2017), "Control Barrier Function Based Quadratic Programs for Safety Critical Systems," IEEE TAC 62(8):3861–3876. paper
  • [ames2019cbfsurvey] Ames, Coogan, Egerstedt, Notomista, Sreenath, Tabuada (2019), "Control Barrier Functions: Theory and Applications," ECC. paper
  • [wangames2017] Wang, Ames, Egerstedt (2017), "Safety Barrier Certificates for Collisions-Free Behaviors in Multirobot Systems," IEEE CDC. The multi-robot CBF that the per-pair budget split in ADR-004 §6.2 generalises. paper
  • [morton2025oscbf] Morton, Pavone (2025), "Oblivious Safety-Critical Control via Operator-Splitting Quadratic Programs," arXiv:2503.17678. OSCBF; CONCERTO's inner per-arm filter (see ADR-004 §5). paper
  • [guerrier2024lcbfsurvey] Guerrier, Fouad, Beltrame (2024), "Learning Control Barrier Functions and their Application in Reinforcement Learning: A Survey," arXiv:2404.16879. paper
  • [lindemann2024safety] Lindemann et al. (2024), "Formal Verification and Control with Conformal Prediction (and the Safety of Learning-Enabled Multi-Agent Systems): A Survey," arXiv:2409.00536. The "Garg/Lindemann" safety survey carried by ADR-004 and ADR-014 tier-2 note 45. paper
  • [ballotta2024aoi] Ballotta, Talak (2024), "Optimal Trade-Offs between Reliability and Freshness for Multi-Agent Decision Making with Age of Information," arXiv:2403.05757. Source for the AoI predictor pattern used in CHAMBER's degradation wrapper (ADR-003, ADR-008). paper
  • [cavorsi2022] Cavorsi, Capelli, Sabattini, Gil (2022), "Multi-Robot Adversarial Resilience using Control Barrier Functions," RSS. Source for the nested-CBF degraded-partner budget template referenced by ADR-006 and ADR-008.

3. Conformal prediction and conformal control

Conformal prediction supplies the distribution-free coverage guarantee that the conformal-slack overlay in ADR-004 §6.1 inherits. Shafer and Vovk (2008) [shafer2008conformal] give the original tutorial; the Angelopoulos and Bates (2023) [angelopoulos2023conformal] Foundations and Trends monograph is the modern textbook reference and the one to recommend to a reader new to the subject.

Huriot and Sibai (2025) [huriotsibai2025] bridge conformal prediction and CBF-based safety. Their Theorem 3 average-loss bound is the theoretical anchor for CONCERTO's conformal overlay and is also the project's biggest open question: whether the bound can be sharpened from an average guarantee to a per-step one (deferred to a follow-up ADR — see ADR-014 §Open questions).

  • [shafer2008conformal] Shafer, Vovk (2008), "A Tutorial on Conformal Prediction," JMLR 9:371–421. paper
  • [angelopoulos2023conformal] Angelopoulos, Bates (2023), "Conformal Prediction: A Gentle Introduction," Foundations and Trends in Machine Learning 16(4):494–591. paper
  • [huriotsibai2025] Huriot, Sibai (2025), "Conformal Control Barrier Functions for Safety under Distribution Shift," arXiv:2409.18862. Theorem 3 (average-loss bound) underwrites the conformal-slack overlay in ADR-004 §6.1. paper

4. Benchmarks, data, and generalization

CHAMBER is a wrapper layer above ManiSkill v3 (Tao et al. 2024) [tao2024maniskill3]; the wrapper-only discipline is captured in ADR-001 and enforced by tests/unit/test_no_private_imports.py. BiGym (Chernyadev et al. 2024) [chernyadev2024bigym], RoCoBench (Mandi et al. 2024) [mandi2024rocobench], and SafeBimanual (Su et al. 2024) [su2024safebimanual] are the closest existing bimanual / multi-arm benchmarks; their gap on the Heterogeneity × Black-box-partner × Safety × Manipulation intersection is the reason CHAMBER exists.

THE COLOSSEUM (Pumacay et al. 2024) [pumacay2024colosseum] is the closest precedent for generalization-axis-perturbation evaluation in manipulation; its 14-axis perturbation matrix is methodologically adjacent to CHAMBER's six-axis heterogeneity sweep, though it focuses on visual / environmental perturbations rather than partner heterogeneity. DROID (Khazatsky et al. 2024) [khazatsky2024droid] is the large-scale in-the-wild manipulation dataset whose 564-scene diversity informs CHAMBER's task-lattice selection.

  • [tao2024maniskill3] Tao, Xiang, Shukla, Qin, Hinrichsen, Yuan, Bao, Lin, Liu, Chan, Gao, Li, Mu, Xiao, Gurha, Nagaswamy Rajesh, Choi, Chen, Huang, Calandra, Chen, Luo, Su (2024), "ManiSkill3: GPU Parallelized Robotics Simulation and Rendering for Generalizable Embodied AI," arXiv:2410.00425. paper
  • [chernyadev2024bigym] Chernyadev, Backshall, Ma, Lu, Seo, James (2024), "BiGym: A Demo-Driven Mobile Bi-Manual Manipulation Benchmark," arXiv:2407.07788. paper
  • [mandi2024rocobench] Mandi, Jain, Song (2024), "RoCo: Dialectic Multi-Robot Collaboration with Large Language Models," IEEE ICRA. paper
  • [su2024safebimanual] Su et al. (2024), "SafeBimanual: Diffusion-based Trajectory Optimization for Safe Bimanual Manipulation," arXiv:2508.18268. paper
  • [pumacay2024colosseum] Pumacay, Singh, Duan, Krishna, Thomason, Fox (2024), "THE COLOSSEUM: A Benchmark for Evaluating Generalization for Robotic Manipulation," arXiv:2402.08191. paper
  • [khazatsky2024droid] Khazatsky, Pertsch, Nair, et al. (2024), "DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset," RSS. paper

5. Reproducibility and statistical evaluation

CHAMBER commits up-front to the reporting discipline these three papers establish. Henderson et al. (2018) [henderson2018matters] is the canonical cautionary tale for deep-RL evaluation: single-seed bars, no confidence intervals, cherry-picked checkpoints, and undocumented hyperparameter sweeps together make published results non-reproducible. evaluation.md lists the specific anti-patterns CHAMBER refuses to fall into.

Agarwal et al. (2021) [agarwal2021precipice] introduce the rliable library and a suite of robust aggregate metrics (IQM, optimality gap, performance profiles) that are now standard in deep-RL reporting; the CHAMBER leaderboard renderer emits these alongside the more traditional mean ± 95% CI. Jordan et al. (2020) [jordan2020evaluating] provide the underlying evaluation-comparison framework — minimum detectable effect size at a given sample size — that justifies the seed counts in ADR-009 §Validation criteria. Wilkinson et al. (2016) [wilkinson2016fair] supplies the FAIR stewardship principles CHAMBER artifacts are released under (see evaluation.md §3.4).

  • [henderson2018matters] Henderson, Islam, Bachman, Pineau, Precup, Meger (2018), "Deep Reinforcement Learning that Matters," AAAI. paper
  • [agarwal2021precipice] Agarwal, Schwarzer, Castro, Courville, Bellemare (2021), "Deep Reinforcement Learning at the Edge of the Statistical Precipice," NeurIPS. Introduces the rliable library. paper
  • [jordan2020evaluating] Jordan, Chandak, Cohen, Zhang, Thomas (2020), "Evaluating the Performance of Reinforcement Learning Algorithms," ICML. paper
  • [wilkinson2016fair] Wilkinson et al. (2016), "The FAIR Guiding Principles for Scientific Data Management and Stewardship," Scientific Data 3:160018. paper