dsh-robotic-harness
Robotic Harness is an embodied-intelligence research plugin suite for DeepSeek Harness, described in the README as an early Demo. It provides rh_* tools and Skills for robot asset/CAD inspection, MuJoCo pick-and-place simulation with fault injection, telemetry diagnostics, data pipelines, experiments, evidence bundles and reports, plus literature search and training workflow scaffolding. Backend-dependent modules return structured 'unavailable' diagnostics when backends are missing, and hardware
Installation
dsh plugin --profile rh-demo add ./packages/dsh-bundleOverview
Robotic Harness is an embodied-intelligence research plugin suite for DeepSeek Harness, described in the README as an early Demo. It provides rh_* tools and Skills for robot asset/CAD inspection, MuJoCo pick-and-place simulation with fault injection, telemetry diagnostics, data pipelines, experiments, evidence bundles and reports, plus literature search and training workflow scaffolding. Backend-dependent modules return structured 'unavailable' diagnostics when backends are missing, and hardware
- README describes ~110 rh_* tools and 27 Skills across assets/CAD, ROS 2, control, vision, models, simulation, telemetry, data, experiments and knowledge.
- Supports URDF/MJCF/SDF inspection, inertia and topology validation, mesh stats, SVG preview, URDF→MJCF conversion, CAD inventory and version compare.
- Provides MuJoCo pick-and-place simulation with six deterministic fault-injection modes, batch benchmarks, replay and sim-vs-real gap reports.
- Offers telemetry analysis, anomaly scans, evidence collection, run comparison and rule-based diagnostics that leave final conclusions to a human.
- Includes data inventory, schema, time-sync, alignment, transforms, splits, de-identification, rosbag conversion, LeRobot export and dataset cards.
- Supports experiment spec/matrix/benchmark/metrics/ablation/report workflows plus evidence bundles with hash manifests and standalone viewers.
- Adds knowledge/memory search and literature search with problem→solution proposals; unreachable network calls return structured 'unavailable' results.
- Training tools are workflow scaffolding: plan, dataset discovery, local job prepare and confirmed remote submit; generated scripts are template placeholders.