Physical grounding

Skillspace link to articulated arm — 7 October 2026

Installing a Skillspace on a simulated arm publishes a one-click local installer, authored placement adapter and native articulated-arm environment. The same controller completes 6/6 changed layouts; no-motion and replayed joint commands with jaws held open each complete 0/6. Three layouts repeat successfully under native network denial. Local product guide and source. DVIDIA metadata provides task identity, not a learned policy; the box proxy, privileged state and collision-excluded arm links bound the result.

Runnable cable environment — 7 October 2026

Offline cable training environment publishes the first practice-to-pack prototype: native CPU cable dynamics, local controller search, frozen simulated evaluation, replay and hashed export. The selected controller and fixed feedback baseline both completed 10/10 held-out episodes; zero-force and release controls completed 0/10. Search did not improve held-out completion. A separate native macOS network-denial check completed 3/3 additional episodes. Runnable archive and instructions. Ideal endpoint attachment and privileged state remain explicit; this is not the retrieval baseline, physical gripping, calibrated rope accuracy or GPU qualification.

Recorded contact fixture — 7 October 2026

The separate native CPU contact fixture records 54 deterministic trials across 18 cases and three repeats. The decision report explains why endpoint agreement does not establish fingertip-force fidelity, and why coarse timesteps can miss collisions. Benchmark source, pinned dependencies, per-trial JSON, an independent AI-assisted code review and a bounded Python-network denial check are published. This fixture does not execute the retrieval baseline below or validate rope materials, GPU training or physical robot transfer.

Question: Which observed states, contact signals or controlled synthetic examples improve a defined task beyond coarse video tags?

Status: scoping · Updated 6 October 2026 · Results: not run

Open a research question · Baseline protocol · Bibliography

Primary sources

• trex2026paper — Dantong Niu et al., T-Rex: Tactile-Reactive Dexterous Manipulation, 15 June 2026.
• trex2026data — T-Rex public dataset card, 2026. Subset size, modalities and preprocessing provenance.
• gr00t2025n1 — NVIDIA (Johan Bjorck et al.), GR00T N1: An Open Foundation Model for Generalist Humanoid Robots, 18 March 2025.
• nvidia2025motion — NVIDIA, Building a Synthetic Motion Generation Pipeline for Humanoid Robot Learning, 18 March 2025.
• nvidia2026blueprint — NVIDIA's synthetic manipulation blueprint, repository accessed 2026; this year is an access date.

What is known and unknown

T-Rex reports 100 collected hours of tactile robot data; its public dataset is approximately 50 hours, not the full training mixture. The card includes RGB, tactile/deformation streams, joints/actions and wrench measurements, and records preprocessing such as imputation trex2026paper, trex2026data. The dataset card carries an MIT license tag; pin the revision and audit the actual artifacts before reuse. These terms do not grant rights to unrelated human videos.

NVIDIA's 11-hour simulation generation and GR00T N1's neural-video generation GPU-hours describe different processes. The vendor article's 40% gain has an unresolved denominator in this review; it is not a DVIDIA result nvidia2025motion, gr00t2025n1. The blueprint's 48 GB A6000 and optional separate 80 GB H100-class workflow are documented workflow requirements, not a universal minimum nvidia2026blueprint.

Unknown: whether richer metadata improves held-out-object retrieval, whether annotation cost is worthwhile, and whether a synthetic intervention transfers outside its controlled setting. This topic starts with retrieval; physical execution requires its own robot evaluation.

Open questions

1. pg-01 — Do temporal state descriptions and measured contact descriptors improve retrieval across held-out objects?
2. pg-02 — Which failures require physical sensing rather than additional RGB labels?
3. pg-03 — Does controlled synthetic augmentation improve held-out simulated completion under equal training budgets?

First small experiment

Audit a capped public T-Rex subset for state-aware retrieval. Inspect metadata first, then double-annotate a small eligible subset. Compare coarse tags with temporal preconditions/effects and, separately, measured contact descriptors. Aim for up to 60 episodes across six primitives only if real coverage and a 2 GB download cap permit it.

Measure Recall@5, state compatibility, disagreement, annotation minutes, decoded bytes, runtime and peak memory. The proposed continuation threshold is a ten-percentage-point Recall@5 gain without lower state compatibility. Use CPU and a fixed simple retrieval method first. Stop if rights, sensor provenance or object-disjoint evaluation are unresolved, or if the cap is reached; reduce the sample transparently rather than downloading the whole dataset.

A later simulator experiment must compare fixed real/teleoperated demonstrations with the same set plus generated trajectories under equal updates and held-out scene seeds. Keep strict completion distinct from partial progress, and appearance generation separately costed. This follow-on is not authorized or run by the baseline.

Milestones

• Primary sources reviewed
• Public-subset and annotation pilot reviewed
• Retrieval baseline executed and artifacts published
• Independent review before simulator expansion

Results

Not run. No data acquisition, annotation, training, robot execution or synthetic-generation result is claimed. Published vendor/paper figures remain attributed to their original settings.