Extend ExPLoRe with reliability-aware vision-tactile-language-action routing

[BryceWDesign]

Description

This PR extends ExPLoRe's learned loss-coupling approach into a vision-tactile-language-action research stack for contact-rich robotic manipulation.

The goal is to preserve ExPLoRe's core idea, learned routing of objectives rather than fixed loss weighting, while testing whether that mechanism can be expanded across multimodal and temporal signals that matter during physical interaction.

What this adds

• Spatiotemporal loss-coupled routing across vision, tactile, force, proprioception, language, action-context, and embodiment signals
• Availability, confidence, freshness, and reliability-aware routing
• Reconstruction, alignment, world-model, action, contact, slip, and feasibility objectives
• Routed action inference and prediction-observation reconciliation
• Modality dropout, noise, delay, drift, temporal shuffle, and force-spike interventions
• Counterfactual action-routing calibration
• Routing faithfulness, utilization, entropy, and mutual-information diagnostics
• Deterministic synthetic contact and slip dynamics
• Monte Carlo action uncertainty estimation
• Independent safety-envelope evaluation
• Constraint-aware candidate selection and finite-difference constraint pressure
• Reproducible experiment evidence, SHA-256 provenance, claim gating, and release-manifest verification
• VTLA configuration, documentation, quickstart examples, tests, and CI

Validation

The current release passes:

• 246 tests
• Python compilation
• Ruff
• Anti-placeholder checks
• Evidence verification
• 166-file SHA-256 release-manifest verification
• Deterministic M1 synthetic mechanism gate

The CI matrix is green across Python 3.10, 3.11, 3.12, and 3.13.

Current evidence

The present evidence authority is intentionally limited to:

M1_SYNTHETIC_MECHANISM

Reproduced synthetic results include:

• Primary specialization delta: 0.901625
• Minimum replicated specialization delta: 0.883409
• Calibrated routing faithfulness: 0.857143
• Faithfulness improvement: 0.678571
• Synthetic smoke-training loss reduction: 84.76%
• Contact-regression invariants: PASS

A negative result is also preserved: coupled routing reduced the routed training objective in all four evaluated synthetic trials, but improved action RMSE over the detached comparison in only one of four trials. The stronger synthetic action-superiority claim therefore remains blocked.

Scope

This PR does not claim real-robot validation, real-dataset performance, sim-to-real transfer, production readiness, or demonstrated superiority on physical manipulation systems.

The intent is to contribute a reproducible, falsifiable extension of ExPLoRe that can serve as a foundation for higher-authority multimodal robotics experiments.

Feedback on the architecture, experimental assumptions, routing formulation, and validation strategy is very welcome.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 64db5b63-11bb-4060-a5c9-215ef307018f

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}