{"body":{"source":"LaTeX source","sections":["Introduction","Overview","Problem Formulation","Encoder and freeze","Broadcast and anchor","Receiver","Route head and executor","From Grid Actions to Velocity Commands","Exploration coverage of the corridor decision","A scalar predictor of learnability","Effect of increased exploration noise","Restoring temporal abstraction","Experiments","Setup","Gridworld (Q1)","Rendered pixels under continuous control (Q1, Q3)","Corruption controls (Q2)","Hardware (Q4)","Related Work","Limitations, Conclusion and Future Work"],"sentences":[],"sentences_withheld":228,"role_counts":[{"section":"Introduction","roles":{"setup":9,"background":8,"other":6,"method":4,"result":5,"interpretation":3}},{"section":"Overview","roles":{"setup":1,"method":10,"other":1}},{"section":"Problem Formulation","roles":{"setup":17}},{"section":"Encoder and freeze","roles":{"setup":1,"method":8}},{"section":"Broadcast and anchor","roles":{"setup":2,"method":1,"interpretation":1}},{"section":"Receiver","roles":{"method":5}},{"section":"Route head and executor","roles":{"method":4,"setup":2,"other":1}},{"section":"From Grid Actions to Velocity Commands","roles":{"other":2,"result":3,"interpretation":2,"method":1}},{"section":"Exploration coverage of the corridor decision","roles":{"method":1,"result":5,"interpretation":3}},{"section":"A scalar predictor of learnability","roles":{"setup":7,"method":3,"limitation":2,"interpretation":2,"result":4}},{"section":"Effect of increased exploration noise","roles":{"interpretation":5,"background":1,"result":3,"limitation":1}},{"section":"Restoring temporal abstraction","roles":{"method":3,"interpretation":1,"result":1}},{"section":"Experiments","roles":{"setup":5}},{"section":"Setup","roles":{"setup":20,"result":2}},{"section":"Gridworld (Q1)","roles":{"setup":1,"result":4,"interpretation":1}},{"section":"Rendered pixels under continuous control (Q1, Q3)","roles":{"result":7,"interpretation":3}},{"section":"Corruption controls (Q2)","roles":{"method":3,"result":4,"interpretation":3}},{"section":"Hardware (Q4)","roles":{"method":2,"result":3,"interpretation":2,"limitation":2}},{"section":"Related Work","roles":{"background":11,"method":2,"interpretation":1}},{"section":"Limitations, Conclusion and Future Work","roles":{"limitation":5,"result":1,"interpretation":3,"future_work":4}}],"withheld_reason":"arXiv's license for this paper lets arXiv distribute its full text; Pith publishes the labels and excerpts and links to arXiv for the paper."},"paper":{"title":"Latent Telepathy: Multi-Robot Communication with Self-Supervised Perceptual Latents","authors":["Howard Wang","Han Zheng","Cathy Wu"],"primary":"cs.RO","arxiv_id":"2609.23269","published":"2026-09-20T01:01:39Z","primary_cat":"cs.RO","submitted_at":"2026-09-20T01:01:39Z","version_compiled":"v1"},"notice":"Every quoted sentence is the authors' text from the arXiv abstract and full text. Labels and probabilities are the outputs of classifier-1.13.0 on the question banks abstract-p1-20260923 + fulltext-p1-20260923, exactly as returned. Measurements are the authors' numbers, located and checked by code. No sentence in this record was written by a language model.","schema":"pith.compiled-record/v0","status":"compiled","results":[{"p":0.97,"kind":"measurement","role":"result","text":"Live, over $34$ fresh hazard placements at $120$–$250$ cm under both lighting states, it decoded $102$ of $102$ decisions ($57$ hazard, $45$ clear) correctly, for $1.000$ balanced accuracy against a pre-registered gate of $0.95$ and a Wilson $95%$ lower bound of $0.964$ over decisions and $0.90$ over placements.","unit":"b17.000.04","hedge":{"p":1.0,"label":"asserted"},"causal_p":0.21,"locations":["Hardware (Q4)"]},{"p":0.96,"kind":"measurement","role":"result","text":"In a gridworld with engineered occlusion, the frozen latent solves an occluded-hazard task in five of five seeds while position and trajectory messages stay at chance in fifteen of fifteen.","unit":"b00.003.03","hedge":{"p":1.0,"label":"asserted"},"causal_p":0.27,"locations":["Introduction"]},{"p":0.96,"kind":"measurement","role":"result","text":"On rendered pixels under continuous velocity commands, it reaches $99.7%$ route optimality against a $49.6%$ no-message floor, matches the oracle, and outperforms the raw image at $1/186$ of the bandwidth.","unit":"b00.003.04","hedge":{"p":1.0,"label":"asserted"},"causal_p":0.33,"locations":["Introduction"]},{"p":0.96,"kind":"observation","role":"result","text":"Corrupting the message at evaluation returns the trained policies to chance.","unit":"b00.003.05","hedge":{"p":1.0,"label":"asserted"},"causal_p":0.91,"locations":["Introduction"]},{"p":0.96,"kind":"measurement","role":"result","text":"This condition also failed, with route optimality of $0.44$ and $0.53$ in two runs against $0.51$ with no message ($n = 2$, indicative only).","unit":"b07.001.04","hedge":{"p":0.95,"label":"asserted"},"causal_p":0.14,"locations":["From Grid Actions to Velocity Commands"]},{"p":0.96,"kind":"measurement","role":"result","text":"The frozen latent solves the task in $5$ of $5$ seeds (success $1.00 ±0.00$), the predicted latent in $4$ of $5$, and the raw $225$-cell observation in $3$ of $5$.","unit":"b14.000.01","hedge":{"p":1.0,"label":"asserted"},"causal_p":0.08,"locations":["Gridworld (Q1)"]},{"p":0.96,"kind":"measurement","role":"result","text":"The latent reaches $0.997 ±0.006$ route optimality against an anchored floor of $0.496 ±0.020$, with no seed overlap ($0.986$–$1.000$ versus $0.472$–$0.522$, $p = 0.004$), and matches the noiseless oracle ($0.988 ±0.006$).","unit":"b15.000.01","hedge":{"p":1.0,"label":"asserted"},"causal_p":0.12,"locations":["Rendered pixels under continuous control (Q1, Q3)"]},{"p":0.96,"kind":"measurement","role":"result","text":"Under zeroed content the head collapses to a constant corridor and episode length rises from $∼209$ to $380$–$440$ steps.","unit":"b16.000.04","hedge":{"p":1.0,"label":"asserted"},"causal_p":0.6,"locations":["Corruption controls (Q2)"]},{"p":0.96,"kind":"measurement","role":"result","text":"On a quarantined validation session at unseen distances under dimmer light it scores $0.929$ balanced accuracy.","unit":"b17.000.03","hedge":{"p":1.0,"label":"asserted"},"causal_p":0.13,"locations":["Hardware (Q4)"]},{"p":0.95,"kind":"measurement","role":"result","text":"We also test the frozen encoder on a physical robot's camera, where it decodes the hazard in $102$ of $102$ live decisions.","unit":"b00.003.06","hedge":{"p":1.0,"label":"asserted"},"causal_p":0.32,"locations":["Introduction"]},{"p":0.95,"kind":"observation","role":"result","text":"Under velocity control, six training configurations that varied the hazard penalty, the discount factor, the spawn curriculum and the arena layout all converged to the same policy, which commits to a single corridor regardless of slab placement (route optimality $≈0.5$).","unit":"b07.001.01","hedge":{"p":1.0,"label":"asserted"},"causal_p":0.39,"locations":["From Grid Actions to Velocity Commands"]},{"p":0.95,"kind":"measurement","role":"result","text":"Task success tracks route optimality within $0.005$ throughout, hazard contacts are zero for the latent and the oracle, and the predicted latent performs identically.","unit":"b15.000.03","hedge":{"p":1.0,"label":"asserted"},"causal_p":0.13,"locations":["Rendered pixels under continuous control (Q1, Q3)"]},{"p":0.95,"kind":"measurement","role":"result","text":"Position and trajectory reach $0.484$ and $0.477$, indistinguishable from the floor.","unit":"b15.000.04","hedge":{"p":1.0,"label":"asserted"},"causal_p":0.06,"locations":["Rendered pixels under continuous control (Q1, Q3)"]},{"p":0.95,"kind":"measurement","role":"result","text":"The raw image, $186×$ wider, reaches $0.819 ±0.230$.","unit":"b15.000.06","hedge":{"p":1.0,"label":"asserted"},"causal_p":0.09,"locations":["Rendered pixels under continuous control (Q1, Q3)"]},{"p":0.95,"kind":"measurement","role":"result","text":"With the wire intact every head scores $1.000$.","unit":"b16.000.02","hedge":{"p":1.0,"label":"asserted"},"causal_p":0.11,"locations":["Corruption controls (Q2)"]}],"abstract":{"source":"arXiv listing abstract v1","sentences":[{"i":0,"p":0.99,"id":"s0","own":false,"role":"background","text":"In a decentralized multi-robot team under partial observability, the fact that decides a robot's next action is often visible only to a teammate.","hedge":{"p":0.81,"label":"asserted"},"own_p":0.09,"main_claim_p":0.03,"quantitative":false,"quantitative_p":0.04,"role_probabilities":{"gap":0.01,"method":0.0,"result":0.0,"background":0.99,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":1,"p":0.58,"id":"s1","own":false,"role":"gap","text":"Existing decentralized methods communicate kinematic information, such as position or planned trajectory, which cannot convey what the teammate perceives.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.05,"main_claim_p":0.02,"quantitative":false,"quantitative_p":0.12,"role_probabilities":{"gap":0.58,"method":0.0,"result":0.0,"background":0.42,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":2,"p":0.5,"id":"s2","own":false,"role":"gap","text":"Learned communication in multi-agent reinforcement learning (MARL) can carry perceptual content, but the resulting messages are task-coupled and opaque.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.05,"main_claim_p":0.03,"quantitative":false,"quantitative_p":0.09,"role_probabilities":{"gap":0.5,"method":0.0,"result":0.0,"background":0.49,"limitation":0.01,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":3,"p":1.0,"id":"s3","own":true,"role":"contribution","text":"We propose Latent Telepathy.","hedge":{"p":0.72,"label":"asserted"},"own_p":0.67,"main_claim_p":0.04,"quantitative":false,"quantitative_p":0.02,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":1.0,"interpretation":0.0}},{"i":4,"p":0.93,"id":"s4","own":true,"role":"method","text":"Each robot broadcasts the perceptual latent vector it already computes for its own use, the output of an encoder trained with a self-supervised joint-embedding predictive objective, frozen, and shared across the team.","hedge":{"p":0.99,"label":"asserted"},"own_p":0.86,"main_claim_p":0.09,"quantitative":false,"quantitative_p":0.09,"role_probabilities":{"gap":0.0,"method":0.93,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.07,"interpretation":0.0}},{"i":5,"p":0.95,"id":"s5","own":true,"role":"method","text":"A teammate learns to act on it from task reward alone.","hedge":{"p":0.99,"label":"asserted"},"own_p":0.51,"main_claim_p":0.1,"quantitative":false,"quantitative_p":0.03,"role_probabilities":{"gap":0.0,"method":0.95,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.05,"interpretation":0.0}},{"i":6,"p":0.57,"id":"s6","own":true,"role":"method","text":"Because the encoder already runs for perception, the message costs no additional computation and a single compact vector of bandwidth.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.64,"main_claim_p":0.05,"quantitative":true,"quantitative_p":0.5,"role_probabilities":{"gap":0.0,"method":0.57,"result":0.03,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.28,"interpretation":0.12}},{"i":7,"p":0.77,"id":"s7","own":true,"role":"method","text":"Because the encoder is frozen before any policy is trained, the message means the same thing to every robot, and the receiving robot is never told what it means.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.57,"main_claim_p":0.07,"quantitative":false,"quantitative_p":0.08,"role_probabilities":{"gap":0.0,"method":0.77,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.09,"interpretation":0.14}},{"i":8,"p":0.97,"id":"s8","own":true,"role":"method","text":"We evaluate Latent Telepathy with a content-controlled protocol in which bandwidth, latency, topology and receiver are held fixed and only the message content varies.","hedge":{"p":0.91,"label":"asserted"},"own_p":0.93,"main_claim_p":0.03,"quantitative":false,"quantitative_p":0.09,"role_probabilities":{"gap":0.0,"method":0.97,"result":0.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.03,"interpretation":0.0}},{"i":9,"p":1.0,"id":"s9","own":true,"role":"result","text":"Broadcasting the latent lets a navigator avoid an occluded hazard in 99.7% of episodes, matching a noiseless hand-designed message.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.96,"main_claim_p":0.42,"quantitative":true,"quantitative_p":0.99,"role_probabilities":{"gap":0.0,"method":0.0,"result":1.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":10,"p":1.0,"id":"s10","own":true,"role":"result","text":"Position and trajectory messages remain at chance, and the raw camera image, 186 times wider, is less reliable than the compressed latent.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.94,"main_claim_p":0.27,"quantitative":true,"quantitative_p":0.98,"role_probabilities":{"gap":0.0,"method":0.0,"result":1.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":11,"p":1.0,"id":"s11","own":true,"role":"result","text":"The result holds from a discrete gridworld to rendered pixels under continuous velocity control, and the encoder decodes the hazard from a physical robot's camera in 102 of 102 live decisions.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.96,"main_claim_p":0.28,"quantitative":true,"quantitative_p":0.99,"role_probabilities":{"gap":0.0,"method":0.0,"result":1.0,"background":0.0,"limitation":0.0,"future_work":0.0,"contribution":0.0,"interpretation":0.0}},{"i":12,"p":0.59,"id":"s12","own":true,"role":"contribution","text":"We also identify a requirement for porting MARL communication results to continuous control, that the decision a message informs must remain reachable by exploration, and show how to restore it.","hedge":{"p":1.0,"label":"asserted"},"own_p":0.93,"main_claim_p":0.11,"quantitative":false,"quantitative_p":0.09,"role_probabilities":{"gap":0.0,"method":0.0,"result":0.19,"background":0.0,"limitation":0.04,"future_work":0.01,"contribution":0.59,"interpretation":0.17}}]},"compiled":{"main_claim":{"p":0.83,"text":"Broadcasting the latent lets a navigator avoid an occluded hazard in 99.7% of episodes, matching a noiseless hand-designed message.","location":"abstract s9","hedge_strength":{"p":1.0,"label":"asserted"}},"provenance":{"route":{"via":"TypeSafe API","model":"classifier-1.13.0","endpoint":"https://api.typesafe.ai/v1/systemone","provider":"TypeSafe"},"packing":"abstract: one request per paper; body: one sentence per request","requests":303,"input_tokens":377764,"cost_usd_reported":0.0,"cost_usd_at_list_price":0.015866},"compiled_at":"2026-09-24","bank_version":"abstract-p1-20260923 + fulltext-p1-20260923","model_version":"classifier-1.13.0","abstract_states":{"claims_first":0.07,"states_limitation":0.81,"human_participants":0.02,"controlled_comparison":0.94,"randomized_assignment":0.02,"states_quantitative_result":0.99,"compares_with_prior_results":0.95,"mentions_code_or_data_availability":0.03},"abstract_vs_body":{"stronger_p":0.67,"supported_p":0.61},"evidential_basis":{"p":0.81,"value":"observation or experiment","probabilities":{"argument":0.0,"literature":0.0,"proof_or_derivation":0.0,"analysis_of_existing_data":0.0,"observation_or_experiment":0.81,"simulation_or_computation":0.19}},"contribution_type":{"p":1.0,"value":"method or tool","probabilities":{"review":0.0,"method_or_tool":1.0,"empirical_study":0.0,"simulation_study":0.0,"theoretical_result":0.0,"dataset_or_benchmark":0.0,"position_or_commentary":0.0}},"contribution_statement":{"p":0.78,"text":"Each robot broadcasts the perceptual latent vector it already computes for its own use, the output of an encoder trained with a self-supervised joint-embedding predictive objective, frozen, and shared across the team.","location":"abstract s4"}},"coverage":{"tier":"full_text","levels_run":["P abstract","S sections","T body sentences","C citations"],"not_examined":["figures and tables","equation content","supplementary material","relations to other papers"],"sources_read":["arXiv listing abstract v1","arXiv full text v1, LaTeX source: 20 sections, 228 body sentences"],"levels_not_run":["R cross-paper relations"],"references_parsed":31,"sentences_in_body":228,"sentences_examined":228,"citation_contexts_examined":48,"published_view":"labels and excerpts (paper license: arXiv distribution only)"},"conditions":[{"p":0.96,"key":"assumption or scope","text":"One scene, one map seed and one encoder seed per substrate.","unit":"b19.000.00","location":"Limitations, Conclusion and Future Work"},{"p":0.95,"key":"assumption or scope","text":"Bandwidth, one-step latency, topology and receiver architecture are held fixed and only the message content varies.","unit":"b00.003.01","location":"Introduction"},{"p":0.94,"key":"assumption or scope","text":"The message function $f$ is fixed before training and identical across robots.","unit":"b02.000.03","location":"Problem Formulation"},{"p":0.94,"key":"assumption or scope","text":"The decisive fact is one bit and a hand-designed oracle ties the latent, so we have shown that a frozen latent is sufficient for reward-only recruitment, not that it is necessary.","unit":"b19.000.02","location":"Limitations, Conclusion and Future Work"},{"p":0.93,"key":"assumption or scope","text":"The scout is stationary.","unit":"b19.000.03","location":"Limitations, Conclusion and Future Work"},{"p":0.92,"key":"assumption or scope","text":"Every message is padded to a common width $D$, carries the same anchor (Section [ref]), travels the same channel with the same latency and reaches the same receiver trained under the same budget.","unit":"b02.002.01","location":"Problem Formulation"},{"p":0.92,"key":"assumption or scope","text":"The pixel result is a contextual bandit over a frozen executor, not end-to-end multi-agent learning, and the executor's route-conditioned pretraining is part of the remedy rather than a neutral substrate.","unit":"b19.000.01","location":"Limitations, Conclusion and Future Work"},{"p":0.91,"key":"assumption or scope","text":"For the comparison to be about content alone, everything else must be held fixed.","unit":"b02.002.00","location":"Problem Formulation"},{"p":0.9,"key":"assumption or scope","text":"The team receives a shared reward $r_t$, and $r_t$ is the only training signal available to the policies.","unit":"b02.000.02","location":"Problem Formulation"},{"p":0.9,"key":"assumption or scope","text":"Because the scout is stationary, its position and trajectory are constant within an episode and carry no information about the hazard, so the comparison measures what perceptual content adds beyond position at matched bandwidth.","unit":"b14.000.05","location":"Gridworld (Q1)"},{"p":0.88,"key":"assumption or scope","text":"At step $t$ robot $i$ receives an egocentric observation $o^i_t$, emits a message $m^i_t = f(o^i_t) ∈ℝ^D$ onto a shared channel, and selects an action $a^i_t ∼π_i( ·|o^i_t, {m^j_t-1}_j ∈𝒩(i))$, where $𝒩(i)$ is the set of robots within communication range and messages arrive with one step of latency.","unit":"b02.000.01","location":"Problem Formulation"},{"p":0.87,"key":"assumption or scope","text":"Under a Gaussian policy over velocity commands the same choice requires roughly thirty consecutive steps of coordinated lateral deviation, which independent per-step noise almost never produces.","unit":"b00.004.03","location":"Introduction"},{"p":0.87,"key":"assumption or scope","text":"Under iid Gaussian exploration the window mean has standard deviation $σ/√(N)$, so the required deviation lies [equation] standard deviations from the mean at the trained $σ$, an event with probability of order $10^-14$ per window.","unit":"b09.000.03","location":"A scalar predictor of learnability"},{"p":0.87,"key":"assumption or scope","text":"The robot was not driven closed-loop.","unit":"b17.000.06","location":"Hardware (Q4)"},{"p":0.85,"key":"assumption or scope","text":"In a gridworld one action moves the agent a whole cell, so choosing the other corridor is a single draw of the policy and every training batch contains both routes.","unit":"b00.004.02","location":"Introduction"},{"p":0.85,"key":"assumption or scope","text":"The boost anneals over this period, so a single run cannot distinguish active suppression from decay, and the ratio above, which treats in- and out-of-window advantages as equal in magnitude, is a heuristic rather than a proof.","unit":"b10.000.08","location":"Effect of increased exploration noise"},{"p":0.85,"key":"assumption or scope","text":"Live, over $34$ fresh hazard placements at $120$–$250$ cm under both lighting states, it decoded $102$ of $102$ decisions ($57$ hazard, $45$ clear) correctly, for $1.000$ balanced accuracy against a pre-registered gate of $0.95$ and a Wilson $95%$ lower bound of $0.964$ over decisions and $0.90$ over placements.","unit":"b17.000.04","location":"Hardware (Q4)"},{"p":0.84,"key":"assumption or scope","text":"Nothing else is trained.","unit":"b01.001.03","location":"Overview"},{"p":0.83,"key":"assumption or scope","text":"The encoder never sees reward, the low-level controller is pretrained and frozen, and the navigator is never told what means.","unit":"b01.001.04","location":"Overview"},{"p":0.83,"key":"assumption or scope","text":"This is a lower bound.","unit":"b09.000.04","location":"A scalar predictor of learnability"}],"measurements":[{"p":0.99,"kind":"percent","unit":"b00.003.04","group":"","value":"99.7%","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"stated once","status":"not repeated"}],"location":"Introduction","quantity":"rendered pixels under continuous velocity commands, it reaches"},{"p":0.99,"kind":"percent","unit":"b00.003.04","group":"","value":"49.6%","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"stated once","status":"not repeated"}],"location":"Introduction","quantity":"commands, it reaches 99.7% route optimality against a"},{"p":0.99,"kind":"percent","unit":"b08.001.01","group":"","value":"100%","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"stated once","status":"not repeated"}],"location":"Exploration coverage of the corridor decision","quantity":"rollouts the policy enters the incumbent corridor in"},{"p":0.91,"kind":"value unit","unit":"b09.000.01","group":"","value":"1.25 m","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"A scalar predictor of learnability","status":"consistent"}],"location":"A scalar predictor of learnability","quantity":"navigator spawns equidistant from the two corridor centers"},{"p":0.98,"kind":"value unit","unit":"b09.000.02","group":"","value":"0.5 m/s","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"A scalar predictor of learnability, Corruption controls (Q2), Effect of increased exploration noise","status":"consistent"}],"location":"A scalar predictor of learnability","quantity":"With v_max ="},{"p":0.98,"kind":"value unit","unit":"b09.000.02","group":"","value":"0.1 s","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"A scalar predictor of learnability, Effect of increased exploration noise","status":"consistent"}],"location":"A scalar predictor of learnability","quantity":"With v_max = 0.5 m/s and a"},{"p":0.98,"kind":"value unit","unit":"b09.000.02","group":"","value":"0.05 m","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"stated once","status":"not repeated"}],"location":"A scalar predictor of learnability","quantity":"of full lateral command displaces the robot by"},{"p":0.84,"kind":"value unit","unit":"b13.000.07","group":"","value":"0.5 m","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"A scalar predictor of learnability, Corruption controls (Q2), Effect of increased exploration noise","status":"consistent"}],"location":"Setup","quantity":"The gridworld has"},{"p":0.91,"kind":"value unit","unit":"b13.000.08","group":"","value":"60 Hz","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"A scalar predictor of learnability, Effect of increased exploration noise, Exploration coverage of the corridor decision","status":"consistent"}],"location":"Setup","quantity":"in a GPU-accelerated physics simulator with physics at"},{"p":0.91,"kind":"value unit","unit":"b13.000.08","group":"","value":"10 Hz","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"A scalar predictor of learnability, Corruption controls (Q2), Effect of increased exploration noise","status":"consistent"}],"location":"Setup","quantity":"simulator with physics at 60 Hz, control at"},{"p":0.91,"kind":"value unit","unit":"b13.000.08","group":"","value":"0.5 m/s","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"A scalar predictor of learnability, Corruption controls (Q2), Effect of increased exploration noise","status":"consistent"}],"location":"Setup","quantity":"velocity commands (v_x, v_y, ω) ∈[-1,1]^3 scaled by"},{"p":0.99,"kind":"percent","unit":"b13.002.01","group":"","value":"93.5%","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"stated once","status":"not repeated"}],"location":"Setup","quantity":"it passed were a linear hazard-visible probe at"},{"p":0.99,"kind":"percent","unit":"b13.002.01","group":"","value":"85.7%","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"stated once","status":"not repeated"}],"location":"Setup","quantity":"a linear hazard-visible probe at 93.5% against an"},{"p":0.97,"kind":"scientific","unit":"b13.002.03","group":"","value":"3×10^-3","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"A scalar predictor of learnability, Corruption controls (Q2), Effect of increased exploration noise","status":"consistent"}],"location":"Setup","quantity":"at step 2, and trains at learning rate"},{"p":0.88,"kind":"scientific","unit":"b13.002.04","group":"","value":"2.5×10^-4","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"stated once","status":"not repeated"}],"location":"Setup","quantity":"potential-based shaping [ref] and an obedience-as-success reward at"},{"p":0.98,"kind":"value uncertainty","unit":"b14.000.01","group":"","value":"1.00 ±0.00","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"Rendered pixels under continuous control (Q1, Q3)","status":"consistent"},{"check":"uncertainty is non-negative","status":"consistent"}],"location":"Gridworld (Q1)","quantity":"the task in 5 of 5 seeds (success"},{"p":0.99,"kind":"p value","unit":"b14.000.04","group":"","value":"p = 0.004","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"Rendered pixels under continuous control (Q1, Q3)","status":"consistent"},{"check":"p-value between 0 and 1","status":"consistent"}],"location":"Gridworld (Q1)","quantity":"has no seed overlap (exact one-sided rank test"},{"p":0.99,"kind":"value uncertainty","unit":"b15.000.01","group":"","value":"0.997 ±0.006","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"stated once","status":"not repeated"},{"check":"uncertainty is non-negative","status":"consistent"}],"location":"Rendered pixels under continuous control (Q1, Q3)","quantity":"The latent reaches"},{"p":0.99,"kind":"value uncertainty","unit":"b15.000.01","group":"","value":"0.496 ±0.020","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"stated once","status":"not repeated"},{"check":"uncertainty is non-negative","status":"consistent"}],"location":"Rendered pixels under continuous control (Q1, Q3)","quantity":"±0.006 route optimality against an anchored floor of"},{"p":0.99,"kind":"value uncertainty","unit":"b15.000.01","group":"","value":"0.988 ±0.006","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"stated once","status":"not repeated"},{"check":"uncertainty is non-negative","status":"consistent"}],"location":"Rendered pixels under continuous control (Q1, Q3)","quantity":"= 0.004), and matches the noiseless oracle"},{"p":0.99,"kind":"p value","unit":"b15.000.01","group":"","value":"p = 0.004","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"Gridworld (Q1)","status":"consistent"},{"check":"p-value between 0 and 1","status":"consistent"}],"location":"Rendered pixels under continuous control (Q1, Q3)","quantity":"±0.020, with no seed overlap (0.986–1.000 versus 0.472–0.522"},{"p":0.99,"kind":"value uncertainty","unit":"b15.000.06","group":"","value":"0.819 ±0.230","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"stated once","status":"not repeated"},{"check":"uncertainty is non-negative","status":"consistent"}],"location":"Rendered pixels under continuous control (Q1, Q3)","quantity":"The raw image, 186× wider, reaches"},{"p":0.98,"kind":"p value","unit":"b15.000.08","group":"","value":"p = 0.375","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"stated once","status":"not repeated"},{"check":"p-value between 0 and 1","status":"consistent"}],"location":"Rendered pixels under continuous control (Q1, Q3)","quantity":"against five the comparison is not significant (Fisher"},{"p":0.99,"kind":"range","unit":"b17.000.04","group":"","value":"120–250 cm","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"stated once","status":"not repeated"},{"check":"range bounds in order","status":"consistent"}],"location":"Hardware (Q4)","quantity":"Live, over 34 fresh hazard placements at"},{"p":0.99,"kind":"percent","unit":"b17.000.04","group":"","value":"95%","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"Rendered pixels under continuous control (Q1, Q3), Setup","status":"consistent"}],"location":"Hardware (Q4)","quantity":"a pre-registered gate of 0.95 and a Wilson"},{"p":0.97,"kind":"range","unit":"b17.000.07","group":"","value":"84–97%","checks":[{"check":"number matches the authors' source text","status":"consistent"},{"check":"same number stated elsewhere in the paper","detail":"stated once","status":"not repeated"},{"check":"range bounds in order","status":"consistent"}],"location":"Hardware (Q4)","quantity":"The matched-optics executor is bang-bang"},{"p":0.99,"kind":"percent","unit":"s9","group":"","value":"99.7%","checks":[{"check":"abstract number also stated in the body","status":"consistent"}],"location":"abstract s9","quantity":"lets a navigator avoid an occluded hazard in"}],"citations_out":{"count":31,"shown":[{"p":0.41,"key":"socialjepa2026","ref":"H. Zhang, Y. Wang, Y. Duan, R. Fu, D. Zhao, S. Fan, S. Cao, W. Guo, and X. Zhou, “Social-JEPA: Emergent geometric isomorphism in independently trained world models,” arXiv:2603.02263, 2026.","unit":"b18.002.02","intent":"method or tool","location":"Related Work","citing_text":"Concurrent CS-JEPA and Social-JEPA broadcast predictive latents between agents and read them with ridge probes on labeled episodes [ref] .","load_bearing_p":0.61},{"p":0.42,"key":"gazzaev2026csjepa","ref":"A.-B. Gazzaev, A. Gavrilov, and S. Muravyov, “One future, every robot: Label-efficient collective-state prediction with decentralized JEPA,” arXiv:2607.28443, 2026.","unit":"b18.002.02","intent":"method or tool","location":"Related Work","citing_text":"Concurrent CS-JEPA and Social-JEPA broadcast predictive latents between agents and read them with ridge probes on labeled episodes [ref] .","load_bearing_p":0.61},{"p":0.98,"key":"bardes2022vicreg","ref":"A. Bardes, J. Ponce, and Y. LeCun, “VICReg: Variance-invariance-covariance regularization for self-supervised learning,” in International Conference on Learning Representations (ICLR), 2022.","unit":"b03.000.03","intent":"method or tool","location":"Encoder and freeze","citing_text":"VICReg variance and covariance terms guard against collapse [ref] , and a class-weighted spatial reconstruction auxiliary, which we found necessary in the gridworld, is carried forward.","load_bearing_p":0.6},{"p":0.44,"key":"zhou2022gnn","ref":"Y. Zhou, J. Xiao, Y. Zhou, and G. Loianno, “Multi-robot collaborative perception with graph neural networks,” IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 2289–2296, 2022.","unit":"b18.001.01","intent":"background","location":"Related Work","citing_text":"Who2com, When2com, V2VNet, Where2comm and the multi-robot GNN perception of Zhou et al. share intermediate features between agents under bandwidth constraints [ref] .","load_bearing_p":0.55},{"p":0.64,"key":"zhou2024dinowm","ref":"G. Zhou, H. Pan, Y. LeCun, and L. Pinto, “DINO-WM: World models on pre-trained visual features enable zero-shot planning,” in International Conference on Machine Learning (ICML), 2025.","unit":"b18.002.01","intent":"supports this paper","location":"Related Work","citing_text":"Joint-embedding predictive encoders and world models produce, without labels, the perceptual latent a teammate would need [ref] .","load_bearing_p":0.54},{"p":0.59,"key":"assran2023ijepa","ref":"M. Assran, Q. Duval, I. Misra, P. Bojanowski, P. Vincent, M. Rabbat, Y. LeCun, and N. Ballas, “Self-supervised learning from images with a joint-embedding predictive architecture,” in IEEE/CVF Conference on Computer Vision and Pattern Recog","unit":"b18.002.01","intent":"supports this paper","location":"Related Work","citing_text":"Joint-embedding predictive encoders and world models produce, without labels, the perceptual latent a teammate would need [ref] .","load_bearing_p":0.54},{"p":0.36,"key":"sukhbaatar2016commnet","ref":"S. Sukhbaatar, A. Szlam, and R. Fergus, “Learning multiagent communication with backpropagation,” in Advances in Neural Information Processing Systems (NeurIPS), 2016.","unit":"b18.000.01","intent":"background","location":"Related Work","citing_text":"Differentiable and reinforced channels learn the message end-to-end from task reward [ref] ; Lin et al. ground the message in an autoencoder trained jointly with the policy [ref] .","load_bearing_p":0.52},{"p":0.5,"key":"wang2020v2vnet","ref":"T.-H. Wang, S. Manivasagam, M. Liang, B. Yang, W. Zeng, and R. Urtasun, “V2VNet: Vehicle-to-vehicle communication for joint perception and prediction,” in European Conference on Computer Vision (ECCV), 2020.","unit":"b18.001.01","intent":"background","location":"Related Work","citing_text":"Who2com, When2com, V2VNet, Where2comm and the multi-robot GNN perception of Zhou et al. share intermediate features between agents under bandwidth constraints [ref] .","load_bearing_p":0.51},{"p":0.36,"key":"lin2021ground","ref":"T. Lin, M. Huh, C. Stauffer, S.-N. Lim, and P. Isola, “Learning to ground multi-agent communication with autoencoders,” in Advances in Neural Information Processing Systems (NeurIPS), 2021.","unit":"b18.000.01","intent":"background","location":"Related Work","citing_text":"Differentiable and reinforced channels learn the message end-to-end from task reward [ref] ; Lin et al. ground the message in an autoencoder trained jointly with the policy [ref] .","load_bearing_p":0.51},{"p":0.54,"key":"liu2020when2com","ref":"Y.-C. Liu, J. Tian, N. Glaser, and Z. Kira, “When2com: Multi-agent perception via communication graph grouping,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020.","unit":"b18.001.01","intent":"background","location":"Related Work","citing_text":"Who2com, When2com, V2VNet, Where2comm and the multi-robot GNN perception of Zhou et al. share intermediate features between agents under bandwidth constraints [ref] .","load_bearing_p":0.5},{"p":0.96,"key":"das2019tarmac","ref":"A. Das, T. Gervet, J. Romoff, D. Batra, D. Parikh, M. Rabbat, and J. Pineau, “TarMAC: Targeted multi-agent communication,” in International Conference on Machine Learning (ICML), 2019.","unit":"b05.000.00","intent":"method or tool","location":"Receiver","citing_text":"The receiver is a single cross-attention layer in the standard attention-fusion form [ref] .","load_bearing_p":0.5},{"p":0.52,"key":"liu2020who2com","ref":"Y.-C. Liu, J. Tian, C.-Y. Ma, N. Glaser, C.-W. Kuo, and Z. Kira, “Who2com: Collaborative perception via learnable handshake communication,” in IEEE International Conference on Robotics and Automation (ICRA), 2020.","unit":"b18.001.01","intent":"background","location":"Related Work","citing_text":"Who2com, When2com, V2VNet, Where2comm and the multi-robot GNN perception of Zhou et al. share intermediate features between agents under bandwidth constraints [ref] .","load_bearing_p":0.49},{"p":0.6,"key":"assran2025vjepa2","ref":"M. Assran et al., “V-JEPA 2: Self-supervised video models enable understanding, prediction and planning,” arXiv:2506.09985, 2025.","unit":"b18.002.01","intent":"supports this paper","location":"Related Work","citing_text":"Joint-embedding predictive encoders and world models produce, without labels, the perceptual latent a teammate would need [ref] .","load_bearing_p":0.49},{"p":0.49,"key":"das2019tarmac","ref":"A. Das, T. Gervet, J. Romoff, D. Batra, D. Parikh, M. Rabbat, and J. Pineau, “TarMAC: Targeted multi-agent communication,” in International Conference on Machine Learning (ICML), 2019.","unit":"b18.000.01","intent":"background","location":"Related Work","citing_text":"Differentiable and reinforced channels learn the message end-to-end from task reward [ref] ; Lin et al. ground the message in an autoencoder trained jointly with the policy [ref] .","load_bearing_p":0.49},{"p":0.46,"key":"assran2023ijepa","ref":"M. Assran, Q. Duval, I. Misra, P. Bojanowski, P. Vincent, M. Rabbat, Y. LeCun, and N. Ballas, “Self-supervised learning from images with a joint-embedding predictive architecture,” in IEEE/CVF Conference on Computer Vision and Pattern Recog","unit":"b00.002.01","intent":"supports this paper","location":"Introduction","citing_text":"Encoders trained with joint-embedding predictive objectives [ref] produce compact latent vectors that capture the structure of a scene, without labels and without reward, and are increasingly used as the perception front end for control [ref] .","load_bearing_p":0.49}]},"limitations_stated":[{"p":1.0,"kind":[],"text":"The scout is stationary.","unit":"b19.000.03","location":"Limitations, Conclusion and Future Work"},{"p":1.0,"kind":[],"text":"The hardware test is decode-only.","unit":"b19.000.04","location":"Limitations, Conclusion and Future Work"},{"p":0.99,"kind":[],"text":"The pixel result is a contextual bandit over a frozen executor, not end-to-end multi-agent learning, and the executor's route-conditioned pretraining is part of the remedy rather than a neutral substrate.","unit":"b19.000.01","location":"Limitations, Conclusion and Future Work"},{"p":0.98,"kind":[],"text":"One scene, one map seed and one encoder seed per substrate.","unit":"b19.000.00","location":"Limitations, Conclusion and Future Work"},{"p":0.97,"kind":[],"text":"The boost anneals over this period, so a single run cannot distinguish active suppression from decay, and the ratio above, which treats in- and out-of-window advantages as equal in magnitude, is a heuristic rather than a proof.","unit":"b10.000.08","location":"Effect of increased exploration noise"},{"p":0.93,"kind":[],"text":"The matched-optics executor is bang-bang ($84$–$97%$ of pre-clamp action components past the rails on held-out simulated frames) and a static bench cannot certify it safe.","unit":"b17.000.07","location":"Hardware (Q4)"},{"p":0.87,"kind":[],"text":"The decisive fact is one bit and a hand-designed oracle ties the latent, so we have shown that a frozen latent is sufficient for reward-only recruitment, not that it is necessary.","unit":"b19.000.02","location":"Limitations, Conclusion and Future Work"},{"p":0.69,"kind":[],"text":"The robot was not driven closed-loop.","unit":"b17.000.06","location":"Hardware (Q4)"},{"p":0.67,"kind":[],"text":"The optimizer is Adam because under SGD the threshold depends on the learning rate (the majority-solved edge moves from $z ≈3.1$ to $5.8$ as the rate increases from $0.01$ to $1$).","unit":"b09.002.00","location":"A scalar predictor of learnability"}],"methods_and_resources":{},"measured_accuracy":{"url":"https://pith.science/accuracy","measured_on":"2026-10-02","method":"two independent model families labelled a fresh random sample blind; the figure is how often the published label matches their consensus","sentence_role_agreement":0.92,"main_claim_agreement":0.908,"typed_citation_agreement":0.933,"typed_citation_agreement_95_interval":[0.917,0.945],"review_comments_accurate":0.758,"review_comments_accurate_95_interval":[0.684,0.82],"review_caught_admitted_error_found_or_related":0.448,"review_comments_judged":{"judged_by":["gpt-5.6-luna","grok-4.7"],"judged_against":"Pith's own text copy of each paper: text extracted from the PDF, the first 90,000 characters","measured_on":"2026-10-04","accurate":113,"decided":149,"sampled_reviews":200,"split_by_lane":false,"by_lane":"https://pith.science/accuracy#by-lane"}}}