Paper board¶
Where each paper and each piece of work stands: backlog (to analyse) → reading (being analysed) → testing (on the rig or the Orin) → review (result written, waiting for me) → done. 180 cards (141 papers, 39 pieces of work: gate runs, paper drafts, audits, robot days), one markdown file each in Shisso-Site-Data; the Orin's work queue commits every status change there, and this page reads it live. Finished papers are in the register.
Build copy of 2026-11-03.
· ·
Backlog 128to analyse
- Paper feasibility: "The small-data curve — when does a JEPA latent stop being kinematics?" — five data scales (45 → 2,900 episodes), the three failed probes + R² per scale (memo only)
work - LeRobot open PRs: ten that need a concrete change, each with a review draft and a ready patch (suggestion blocks + `git am` patch so the fix carries Ravindu's authorship); contribution opportunities as the second output — nothing posted
work - Paper 4 (own-predictor monitor): no-go for now — the remaining 40 sealed §29 layouts on a free arm day, then revisit
work - RoboSeed: Multidimensional Evaluation of Vision-Language-Action and World-Action Models: Capability, Resource Affordability, and Research Evidence
arXiv:2610.40026 - VLA start suppression
arXiv:2610.41609 - CuraEval
arXiv:2610.38221 - SO-101 demo scaling
arXiv:2610.37170 - SO-101 Bench
arXiv:2610.36420 - VLA-HWBench
arXiv:2610.34717 - exec-logs judges
arXiv:2610.34151 - RoboJev
arXiv:2610.33718 - WorldFailBench
arXiv:2610.33565 - Planning Limits of JEPA
arXiv:2609.39235 - PSG-JEPA + SA/IDM
arXiv:2609.03565 - PSG-JEPA + SA/IDM
arXiv:2608.06799 - SEVO
arXiv:2605.11114 - PEEK
arXiv:2509.18282 - Learning Safe Flow Matching Robot Policies from Binary Feasibility Oracles
arXiv:2610.40075 - SaiVLA-JEPA: Evaluating Future Latent Prediction in JEPA-Conditioned VLA Control
arXiv:2610.38508 - SuperMimic: Scaling Humanoid Whole-Body Control with One Differentiable World Model
arXiv:2610.28480 - LaLa-World: Scaling Latent World Models with Natural Language Actions
arXiv:2610.27851 - EgoXperience: Learning World-Grounded Egocentric Representations from Human Experience
arXiv:2610.26625 - When Better Policies Leave Harder Failures: Coverage Shift in VLM Success Monitoring
arXiv:2610.24801 - Blind Precise Assembly: Proprioceptive Sim-to-Real Peg-in-Hole for Diverse Parts
arXiv:2610.23849 - HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding Agents
arXiv:2610.23479 - JevAgentHarness: Benchmarking, Deploying and Training Decision Models That Share Control with LLMs
arXiv:2610.21412 - Cracking Mobile Manipulation: From Failure Diagnosis to Agentic Robot System
arXiv:2610.19600 - One-Step Prediction Fidelity Is a Poor Proxy for Planning Performance in JEPA World Models
arXiv:2610.19375 - LoopWAM: Closed-Loop Generalist Policy Improvement from Deployment Rollouts
arXiv:2610.18873 - What Matters in Quality-Aware VLA Pretraining: A Benchmark and Empirical Study
arXiv:2610.18869 - SEAR: Self-Evolving Agentic Robotics
arXiv:2610.18792 - ReTrace: Cross-Stage Failure Attribution for Self-Improving Robotic Manipulation
arXiv:2610.15814 - Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation
arXiv:2610.14162 - Embodied Turing Machines: Stateful Code for Robot Recursive Self-Improvement
arXiv:2610.12369 - Long-WAM: Scaling the Context of World-Action Models
arXiv:2610.10528 - Co-Evolving Robot Orchestrators and Policies through Deployment
arXiv:2610.09228 - Harness Engineering for Software Engineering via Modular Executable Dev-Primitives
arXiv:2610.07832 - How (and How Not) to Use Data Augmentation in VLA Post-Training
arXiv:2610.05994 - Mulligan: Performance-Guided Data Collection for Efficient On-Robot Learning
arXiv:2610.05882 - CF-JEPA: Improving Robustness of JEPA World Models via Controllability Factorization
arXiv:2610.00727 - DeepJEPA: Scaling World Models from Within
arXiv:2610.00368 - Ego4WAM: What Matters When Scaling Egocentric Human Data for Robot Learning?
arXiv:2609.40341 - MotorMind: Scaffolding General Vision Language Models for Zero-Shot Robot Manipulation
arXiv:2609.38078 - Spotter: Let the Embodied Model Lead, and the VLM Reflect for It
arXiv:2609.36808 - Self-Evolving Coding Agents: From Digital Programs to Physical-World Intelligence
arXiv:2609.35432 - AD-E2E-JEPA: A Joint-Embedding Predictive Architecture For End-to-End Autonomous Driving
arXiv:2609.34085 - Beyond One-Step Accuracy: State-Affine Latent Transition for Reliable Visual Planning
arXiv:2609.33595 - Not All Errors Matter: Decision-Relevant Prediction Error Predicts Planning Quality
arXiv:2609.32322 - InternW0-$Δ$: A World Action Model Bridging Predictive Dynamics and Actions with 20K+ Hours of Open Data
arXiv:2609.31394 - Jev in the Wild: A Data-Driven Analysis of the Jev Model's Functionality, Applications and Ecosystem
arXiv:2609.30216 - RegenHarness: A Robot Agent Harness with Evidence-Gated Recursive Self-Improvement
arXiv:2609.27612 - JEV-as-a-Judge: Accept When Confident, Escalate When Unsure
arXiv:2609.26550 - Workspace Models: Lightweight Robotic Memory via Saliency-Driven Supervision
arXiv:2609.20820 - From Rollout to Reset: A Graph-Based Harness for Autonomous Long-Horizon Manipulation Evaluation
arXiv:2609.19413 - KINO: A Keyframe Interface for VLM Planning and Whole-Body Control in Humanoid Loco-Manipulation
arXiv:2609.18869 - Reinforcement Learning for Real-Time Vision-Language-Action Policies
arXiv:2609.18207 - The Robot Data Factory
arXiv:2609.16705 - Breaking the Token Ceiling: Distilling Smaller, Stronger Byte Models
arXiv:2609.12303 - GAUGE: When Not to Trust LLM-as-a-Judge in User-Simulated Evaluation of Task-Oriented Agents
arXiv:2609.12191 - FARM: Reading Failure Signals from the Internal Predictive States of a Frozen Robotic World Model
arXiv:2609.11445 - Show-Harness: Just a VLM Agent Can Play Robots
arXiv:2609.10522 - No Free Checker: A Survey of Verifiers for Robot Policies
arXiv:2609.09250 - SMELT: Scaling Laws for Compute-Matched MoE Looped Transformers
arXiv:2609.01343 - Puppeteer: Object-Grounded Posture-Aware Co-Speech Gesture Generation
arXiv:2609.00369 - Puro-2B: Poor Lab's Qwen2-1.5B Trained on RTX 5090 within $5090
arXiv:2608.27370 - Zero-WAM: In-Context World-Action Modeling from Human Videos for Open-Ended Task Generalization
arXiv:2608.26103 - StreamPI: Streaming Multimodal Temporal Modeling for Vision-Language-Action Models
arXiv:2608.26067 - Learning to Act While Waiting: RL Finetuning of Generalist Robot Policies Under Inference Latency
arXiv:2608.23831 - Zetta $ζ$: An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence
arXiv:2608.16590 - Skills in Weights, Memory in Code: Hybrid Learning for Memory-Dependent Robot Manipulation
arXiv:2608.09410 - JEPA-WAM: Learning Vision-Language-Action Policies with Joint-Embedding World Modeling
arXiv:2608.09381 - Ego2Robot: Scalable Robot Data Synthesis from Egocentric Human Data
arXiv:2608.02580 - Temporal-Distance JEPA: Plan-Aware Representation Learning for Latent World Model Predictive Control
arXiv:2607.25337 - ArmnetBench v0.1: Parallel Real-World Evaluation of Manipulation Policies on a Low-Cost Arm Farm
arXiv:2607.24481 - Addressing the Orchestration Gap in Generalist Robots via Physical Agency
arXiv:2607.21725 - ABot-World-0: Infinite Interactive World Rollout on a Single Desktop GPU
arXiv:2607.19191 - Patch Policy: Efficient Embodied Control via Dense Visual Representations
arXiv:2607.18236 - Foresight Residual RL for Long-Horizon Robot Manipulation with Vision-Language-Action Models
arXiv:2607.16506 - The JEPA Predictor: A Transferable Operator for Occluded Feature Completion
arXiv:2607.16274 - Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories
arXiv:2607.15330 - RoboTTT: Context Scaling for Robot Policies
arXiv:2607.15275 - GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by AutoResearch
arXiv:2607.13960 - WAM-TTT: Steering World-Action Models by Watching Human Play at Test Time
arXiv:2607.06988 - Optimal Transport Q-Learning for Flow Policy Steering and Acceleration
arXiv:2607.06262 - RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies
arXiv:2607.04434 - GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation
arXiv:2607.02642 - AgentLTL: A Trace-Verification Framework for Measuring, Enforcing, and Training Procedural Compliance in Tool-Using LLM Agents
arXiv:2607.02599 - Stage-Transition Dense Reward Modeling for Reinforcement Learning
arXiv:2606.31377 - PhysisForcing: Physics Reinforced World Simulator for Robotic Manipulation
arXiv:2606.28128 - Fast LeWorldModel
arXiv:2606.26217 - HiL-ResRL: A Model-Agnostic Finetuning Adapter via Human-in-the-loop Residual Reinforcement Learning
arXiv:2606.22860 - ENPIRE: Agentic Robot Policy Self-Improvement in the Real World
arXiv:2606.19980 - ROBOSHACKLES: A Safety Dataset for Human-Injury Prevention in Embodied Foundation Models
arXiv:2606.18632 - SC3-Eval: Evaluating Robot Foundation Models via Self-Consistent Video Generation
arXiv:2606.18610 - DREAM-Chunk: Reactive Action Chunking with Latent World Model
arXiv:2606.18589 - FACTR 2: Learning External Force Sensing for Commodity Robot Arms Improves Policy Learning
arXiv:2606.12406 - SARM2: Multi-Task Stage Aware Reward Modeling for Self Improving Robotic Manipulation
arXiv:2606.10305 - Efficient-WAM: A 1B-Parameter World-Action Model with Low-Cost Future Imagination
arXiv:2606.10040 - FF-JEPA: Long-Horizon Planning in World Models with Latent Planners
arXiv:2606.09311 - Benchmarking Vision-Language-Action Models on SO-101: Failure and Recovery Analysis
arXiv:2606.08881 - EXPO-FT: Sample-Efficient Reinforcement Learning Finetuning for Vision-Language-Action Models
arXiv:2605.25477 - Towards Long-horizon Embodied Agents with Tool-Aligned Vision-Language-Action Models
arXiv:2605.13119 - MolmoAct2: Action Reasoning Models for Real-world Deployment
arXiv:2605.02881 - $π_{0.7}$: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities
arXiv:2604.15483 - Generalizable Dense Reward for Long-Horizon Robotic Tasks
arXiv:2604.00055 - V-JEPA 2.1: Unlocking Dense Features in Video Self-Supervised Learning
arXiv:2603.14482 - HomeGuard: VLM-based Embodied Safeguard for Identifying Contextual Risk in Household Task
arXiv:2603.14367 - vla-eval: A Unified Evaluation Harness for Vision-Language-Action Models
arXiv:2603.13966 - Quality over Quantity: Demonstration Curation via Influence Functions for Data-Centric Robot Learning
arXiv:2603.09056 - TOPReward: Token Probabilities as Hidden Zero-Shot Rewards for Robotics
arXiv:2602.19313 - Scaling Verification Can Be More Effective than Scaling Policy Learning for Vision-Language-Action Alignment
arXiv:2602.12281 - Causal World Modeling for Robot Control
arXiv:2601.21998 - RoboSafe: Safeguarding Embodied Agents via Executable Safety Logic
arXiv:2512.21220 - Towards Accessible Physical AI: LoRA-Based Fine-Tuning of VLA Models for Real-World Robot Control
arXiv:2512.11921 - MIND-V: Hierarchical World Model for Long-Horizon Robotic Manipulation with RL-based Physical Alignment
arXiv:2512.06628 - LeJEPA: Provable and Scalable Self-Supervised Learning Without the Heuristics
arXiv:2511.08544 - RoboHiMan: A Hierarchical Evaluation Paradigm for Compositional Generalization in Long-Horizon Manipulation
arXiv:2510.13149 - FailSafe: Reasoning and Recovery from Failures in Vision-Language-Action Models
arXiv:2510.01642 - SARM: Stage-Aware Reward Modeling for Long Horizon Robot Manipulation
arXiv:2509.25358 - LocoFormer: Generalist Locomotion via Long-context Adaptation
arXiv:2509.23745 - OpenGVL -- Benchmarking Visual Temporal Progress for Data Curation
arXiv:2509.17321 - Robot Control Stack: A Lean Ecosystem for Robot Learning at Scale
arXiv:2509.14932 - CUPID: Curating Data your Robot Loves with Influence Functions
arXiv:2506.19121 - V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning
arXiv:2506.09985 - SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learning
arXiv:2503.03480 - MuJoCo Playground
arXiv:2502.08844 - Vision Language Models are In-Context Value Learners
arXiv:2411.04549 - Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware
arXiv:2304.13705
Reading 4being analysed
- Shisso paper JEPA: a small JEPA world model on a low-cost arm
work - Shisso paper A: reproducibility of the papers register
work - Shisso paper B: failure monitors and the episode-length null for robot reward models
work - Shisso paper C: a sync-mode inference recipe for a VLA on an edge GPU
work
Testing 1on the rig or the Orin
- Smooth Action Chunking Via Movement-Primitive Representations for Imitation Learning
arXiv:2610.39451
Review 1result written, waiting for the owner
- TurboVLA + transformers 5 upstream drafts
work
Done 46
- Robometer fail
arXiv:2603.02115 · noteA 4B reward model judges SO-101 rollouts zero-shot.
- TurboVLA withdrawn
arXiv:2607.27205 · noteOur claim that the architecture made the policy ignore its instruction.
- VLA-Corrector fail
arXiv:2609.06508 · noteA verifier on observable history flags a failing rollout early enough to correct it.
- REKI reproduces
arXiv:2610.22178 · notePolicies without memory oscillate between poses at a measurable rate.
- HWM (hierarchical latent world models) not proven
arXiv:2604.03208 · noteA high-level world model over learned macro-actions beats the low-level model beyond about 1 s (their Fig. 6 crossover).
- Adaptive Action Chunking (AAC) fail
arXiv:2604.04161 · noteChunk length chosen from the entropy of N sampled chunks follows the task's phases (long in transport, short at the grasp), lowers the boundary jump, and costs little with batched sampling.
- LeJEPA identifiability reproduces
arXiv:2605.26379 · noteA SIGReg JEPA trained on goal-directed data recovers the true latent state only partially and unevenly.
- VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon
arXiv:2607.01804 - Beyond State-as-Action: Exploiting Command-State Discrepancy for Robot Imitation Learning
arXiv:2609.33145 - Taming VLAs under Robot Execution Errors: Self-Compensation and Stress Testing
arXiv:2609.37334 - When to Switch: Reliable Action-Chunk Extension for Vision-Language-Action Models
arXiv:2610.05719 - Not All Actions Contribute Equally: Learning Precise Progress Estimation from Imprecise Demonstrations
arXiv:2610.23769 - PACT-VLA fail
arXiv:2610.28019 · notePicking among N sampled chunks by their flow residual under the pre-commit observation fits what happens next better than a random or stale pick, with no added control-step delay.
- Horizon × depth (Euler steps vs execution horizon) not proven
arXiv:2610.32102 · noteWith long open-loop commitment, cutting a flow-matching VLA's Euler steps from 10 to 2 contracts the action cloud around an unchanged centroid, saves most of the inference time and avoids edge-GPU thermal throttling.
- REMORA fail
arXiv:2610.39125 · noteReading a progress distribution out at the supported peak nearest the previous estimate steadies the estimate and lets it fall during real setbacks, with no worse detection of the stages.
- §28 action sensitivity of W1 / W2′
work - §30 inverse-dynamics latent-action probe
work - §29 VLA-JEPA own surprise signal: feasibility, then exploratory replay
work - TurboVLA: horizon-fair probe, plugin v0.3.0 on the Orin, H=50 run prepared (not launched)
work - §26 world-model report for W2′
work - VLA-JEPA rotary fix (#4835): environment and measurement checks
work - LeRobot PR reviews: Orin numbers (GPU session)
work - Torque-aware VLA-JEPA: Phase 0 recon
work - VLA-JEPA rotary fix: the comparison and the upstream draft
work - §29 step 2: VLA-JEPA's own surprise on the robot (robot day)
work - LeRobot PR reviews batch 3: Orin numbers (offline)
work - My LeRobot PR: processor step settings across save/load, final checks
work - REKI SO-101: frozen-VLM fine-tune on vast (3,500 steps, batch 30 × 4)
work - REKI SO-101: serve on vast, 20-trial eval per task, PR #7 results draft
work - Deep-dive post: VLA chunk-boundary jumps on the Orin, fixed in 0.4 ms (blog + social drafts)
work · note - Research-engineer post series: mine the repo for stories, write up to 5 full posts
work - vla.cpp PR: SmolVLA prefix overwrite (prev_actions, inference_delay), answers #46 — prepared, owner opens
work - Fix the published chunk-boundary post: add LeRobot identity / exact RTC as baselines
work - paper-vs-hardware audit fixes (L0 recomputes, cross-refs, units) + shissorobotics.com phase 3 live
work - Book: verify the 10-27 Levine pass against the register (§16.1c table, 16.2 clause, 16.5a, 16.7b, 8.7), extend it to chapters 6 / 8 / 15 / F, build the interview rehearsal set
work - SURVEYOR (NeurIPS 2026 submission): τ calibration on real-robot LeWM latents, the accept rule as a seventh monitor
work - Claims in the register: per paper its goal and hypotheses, one triple per tested claim (subject / relation / object / metric / conditions), contradictions found at build
work · note - Audit the claims register against the PDFs: every goal, hypothesis, `stated_in`, `theirs`, relation and condition re-read from the paper — citation fixes on a branch, content rows for the owner
work - Paper feasibility 2 — "Episode length as the null model for robot reward models": the short-paper version (memo only)
work - Paper feasibility 4 — "The policy's own predictor as its failure monitor": zero extra parameters (memo only)
work - TOPReward vs elapsed time (WS-081 A4): Qwen3-VL-8B TOPReward on ArmnetBench 2,047 + own rollouts through LENGTH_PROTOCOL.md's analyses, folded into the WS-077 draft
work - Execution on a $200 arm (WS-082 C1 + C2): servo dead zone and command-vs-state gap, leader vs follower side per metric, citing 2609.37334 and 2609.33145 — frozen protocol, offline
work · arXiv:2609.37334, arXiv:2609.33145 - Gripper timing at transitions: the 32× gripper error of WS-086, aligned on G15 stage labels and the STS3215 command floor (G15-v3), offline — feeds smooth-chunking gate B
work - Paper feasibility: "A small JEPA world model learns the arm, not the task" — twelve pre-registered uses of W1 on a $200 robot; step 0 = the checkpoint-load audit of every W1 / W2 gate
work - Paper feasibility: "What a 13 ms latent surprise can see" — detection per perturbation class (displacement, lighting, jitter, frame drop, wrong action, slip, timing) for W1, W2′ and VLA-JEPA's own predictor, at a fixed false-alarm rate (memo only)
work - Score the frozen receptive-field protocol (WS-108): perturbation library, checks, M1/M2 then M3 per class × step, sanity bar, N rows
work