Skip to content

TOWER 1.0.0 release notes

Status: Draft — unreleased; includes the proposed scene revision. Final docs, package versions/dependency bounds, runtime completion and release validation must be reconciled with the final code. This is not a release announcement.

Planned software release

  • Three distributions: nero-kinematics, tower-eval, and tower-sim.
  • One canonical task registry and scoring pipeline, with explicit paper-v1 and bench-2026.11 rulesets and compatible historical task/schema aliases.
  • Shared Nero FK/IK, versioned calibration and command bounds, and a simulator control/evidence implementation developed against the frozen twelve-task suite.
  • XPolicyLab client integration and in-process policies. TOWER policy serving and its vendored policy servers have been removed.
  • Isaac 6.1 / benchmark 2026.11 is the supported simulator line.

The Unreleased changelog records merged changes. The public runtime uses tower_sim.bench and defaults to benchmark 2026.11.1. The public data is published in TOWER-Assets at commit d78e4ac0b2bfc1e10a15d31ac7e7d723be2764f1, tagged benchmark-2026.11.1. Software acceptance remains subject to its GPU, import, posture and exact gates.

Version and evidence boundaries

Packages use SemVer and target 1.0.0. The public benchmark revision 2026.11.1 uses tower-scene-v2, neutral scene identifiers, pinned external evaluation payloads and a committed all-600 physical equivalence map. It has identical resolved physics to the internal predecessor 2026.11.0. The public repository starts with fresh history and contains no predecessor data or tag. Archival compatibility checks require separately supplied evidence and otherwise skip. The public asset, scene and golden locators pin that immutable HF commit. Internal lineage stays outside that dataset. Wire schema versions are independent; old jenga_* strings remain readable.

Historical startup checks, policy diagnostics and command replay comparisons have different scopes. None alone establishes new posture acceptance or a fresh learned-policy benchmark result. Maintainer posture acceptance requires the current no-reversed-grasp rule, task-specific contact geometry, planned approach/carry checks and inspection of the recorded approach. Historical pass receipts do not waive those checks.

Proposed release attachments

The final public tree includes redacted evaluation metadata and a concise validation summary. Attach only revision-labelled, redacted acceptance receipts and their hashes. Internal archival evidence is not a release attachment. The package workflow attaches distributions, the asset lock and checksums.

The asset check confirms the existing public HF tag/commit/manifest pin; it does not upload or qualify model weights.

Scene-v2 evaluation states use the episode wire key and retain physical measured asset identities. Public data excludes archival lineage and source-workflow metadata. All-scene equivalence compares complete runtime inputs against the explicit archive field-removal contract. Data downloading verifies the upload MANIFEST; HF locator pinning is independent of data identity.

Public benchmark runs

Added full-task and subset benchmark runs, sequential batching and shard aggregation without mandatory pilot review. Automatic recorded TCP posture flags are informational. Maintainer qualification remains fail-closed; benchmark revision, scoring rules, data hashes and package versions are unchanged.