Upload 3 files

README.md

@@ -0,0 +1,72 @@
+---
+tags:
+- microduck
+- robotics
+- onnx
+- microduck-slot:walk
+library_name: microduck
+pipeline_tag: robotics
+---
+
+# dfreq-rollers
+
+**Published on Hugging Face.** Repository: `djduecaster/dfreq-rollers`; proposed Arena entry: `dfreq-rollers`. Not submitted to Arena.
+
+A small, event-specific command calibration of Pollen Robotics’ Roller Sprint reference feedback policy. The local full official sweep scored **3.112 seconds, 5/5 finishes**, versus **3.136 seconds** for the unchanged reference: **0.024 seconds (about 0.77%) faster**. The selected five-start result was repeated exactly. This is not a server-verified score or leaderboard-record claim.
+
+## Credit and actual contribution
+
+The base is [Pollen’s Roller Sprint reference](https://huggingface.co/pollen-robotics/microduck-roller-sprint-2m-reference/tree/c72fb92594e8b415487e87dfeed509141afc5e25), revision `c72fb92594e8b415487e87dfeed509141afc5e25`.
+
+- Base SHA-256: `01edf63a529a8396d46dac62bb427fda1ce67685611e72ddf790bf46f610c82a`.
+- Candidate `policy.onnx`: **794,263 bytes**, SHA-256 `2092a1d8e5b30f2778f227ddf6f26f8eda6487236279f8663bdff71e12805590`.
+- **All original neural weights and normalizer bytes are unchanged. No new training or fine-tuning occurred.** The manifest’s `training` field describes Pollen’s upstream training, not work performed for this calibration.
+
+Exactly one predeclared offset was tested. Before the original normalizer, only raw forward-command observation component 48 (zero-based) is mapped:
+
+```text
+inside = (vx_external >= float32(0.3)) AND (vx_external <= float32(2.0))
+vx_internal = Where(inside,
+    Clip(float32(vx_external + float32(-0.05)), float32(0.3), float32(2.0)),
+    vx_external)
+```
+
+The prefix contains only standard ONNX Split/comparison/And/Add/Clip/Where/Concat operations. Removing it and reversing its single input rewire restores the original model bytes exactly. The other 60 observations—including genuine actual previous-action feedback, lateral command and heading error—are unchanged. There is no clock, step counter, seed classifier/table, runtime metadata conditioning or prerecorded action sequence. This remains neural state feedback with a static command calibration, not a newly trained network.
+
+## Interface and command semantics
+
+Model API 1: **61 raw float32 observations → 14 float32 actions at 50 Hz**, **rollers** rig, embedded normalizer, perpetual `walk` slot and standing entry pose. The upstream twist convention is `[push: 0 coast, >0 push m/s, <0 brake; lateral 0; heading error rad]`; the Arena commands `[vx,0,0]`. Head/body commands remain unused zeros.
+
+**Zero/coast, signed zero, negative/brake and positive commands outside the inclusive calibrated interval are returned bit-exactly unchanged.** Positive non-grid values *inside* the interval are calibrated too. This preserves the upstream input modes, not a new stopping/braking or real-robot safety guarantee.
+
+The fixed mask creates a discontinuity just above the upper bound: float32(2.0) maps to approximately 1.95, while the next representable larger input is unchanged. No smoothness claim is made. At the selected external 1.1 command, the exact internal float32 value is **1.0500000715255737**, not decimal1.05.
+
+**Simulation-only and event-specific.** No physical-robot deployment, all-speed stability or general-purpose controller safety has been established. No robot-loading recommendation is provided.
+
+## Local verification
+
+The unchanged Arena runner v4 performed the complete official 18-command scan on seed 1, then the three selected commands across the remaining four starts—**30 official races**. It selected external **1.1 m/s**:
+
+| Start seed | Finish time (s) |
+| --- | ---: |
+| 1 | 3.10 |
+| 2 | 3.08 |
+| 3 | 3.16 |
+| 4 | 3.08 |
+| 5 | 3.14 |
+
+The separate five-race confirmation reproduced all 20 observation/action/timed-trace/untimed-settle arrays exactly. Before the candidate screen, the unchanged reference also matched its 20 cached baseline arrays. Relative to the reference, three starts improved, one tied and one became slower; the aggregate improved from 784 to 778 timed control ticks.
+
+The bounded investigation used **one candidate export, 1818 saved-input preflight evaluations, 45 native races and 19,514 controls**, including settling. The preflight compared 909 candidate outputs byte-for-byte against the unchanged base on independently mapped observations. Source/math review and an independent saved-output audit passed. No physics or neural weights were changed.
+
+These original five starts were used for screening and selection: **they are optimization data, not held-out robustness evidence**. Other commands included falls and timeouts. Exact local repetition does not guarantee identical server results or future eligibility.
+
+## Upstream training provenance
+
+Pollen’s manifest identifies task `Mjlab-RollerSprint2m-MicroDuck`, repository `pollen-robotics/microduck-challenges`, commit `611b6453c`, checkpoint 2999, seed 1, 4096 environments and 3000 iterations, using `mjlab-microduck 0.1.0 @ 981a279c6`. Those are **upstream training facts**; no training was performed for `dfreq-rollers`.
+
+## Rights and publication status
+
+Pollen Robotics’ upstream weights are credited. No standalone open-source model license is asserted or invented. This publication is an attributed Hugging Face service-scoped remix under applicable platform terms, not a blanket off-platform redistribution grant. Publication was authorized separately by the user.
+
+This repository has been published with user authorization. No Arena entry has been submitted.

manifest.json

@@ -0,0 +1,129 @@
+{
+  "schema_version": 2,
+  "model_api": 1,
+  "obs_len": 61,
+  "action_len": 14,
+  "robot": {
+    "model": "microduck",
+    "hw_rev": 1,
+    "servos": "xl330",
+    "control_hz": 50,
+    "accessories": [
+      "rollers"
+    ]
+  },
+  "name": "dfreq-rollers",
+  "kind": "perpetual",
+  "entry_pose": "standing",
+  "description": "Simulation-only Roller Sprint command calibration of unchanged Pollen reference neural weights. A fixed -0.05 m/s offset/clamp applies only inside positive F32 [0.3, 2.0]; zero/coast, negative/brake and commands outside the interval pass through unchanged. Local full official sweep and exact repeat: 3.112 s, 5/5. No new neural training or public score claim.",
+  "command": {
+    "encoding": "constant",
+    "idle": [
+      0.0,
+      0.0,
+      0.0
+    ],
+    "twist": "[push: 0 coast, >0 push m/s, <0 brake; lateral 0; heading error rad]; the Arena commands [vx, 0, 0]",
+    "head": "unused (zeros)",
+    "body": "unused (zeros)"
+  },
+  "duration_s": null,
+  "slot": "walk",
+  "training": {
+    "repo": "https://github.com/pollen-robotics/microduck-challenges.git",
+    "commit": "611b6453c",
+    "branch": "main",
+    "dirty": false,
+    "command": [
+      "train",
+      "Mjlab-RollerSprint2m-MicroDuck",
+      "--env.scene.num-envs",
+      "4096",
+      "--agent.max-iterations",
+      "3000",
+      "--agent.seed",
+      "1",
+      "--agent.logger",
+      "tensorboard",
+      "--agent.run-name",
+      "reference"
+    ],
+    "seed": 1,
+    "base": "mjlab-microduck 0.1.0 @ 981a279c6",
+    "started": "2026-09-26T04:21:50Z",
+    "task_id": "Mjlab-RollerSprint2m-MicroDuck",
+    "checkpoint": 2999,
+    "source_file": "model_2999.pt"
+  },
+  "arena": {
+    "event": "roller-sprint-2m"
+  },
+  "artifact": {
+    "file": "policy.onnx",
+    "bytes": 794263,
+    "sha256": "2092a1d8e5b30f2778f227ddf6f26f8eda6487236279f8663bdff71e12805590"
+  },
+  "provenance": {
+    "base_model": "pollen-robotics/microduck-roller-sprint-2m-reference",
+    "base_revision": "c72fb92594e8b415487e87dfeed509141afc5e25",
+    "base_sha256": "01edf63a529a8396d46dac62bb427fda1ce67685611e72ddf790bf46f610c82a",
+    "original_neural_parameters_byte_exact": true,
+    "training_field_describes_upstream_Pollen_model": true,
+    "new_neural_training": false,
+    "original_contribution": "One predeclared static, memoryless positive-domain velocity-command calibration; no learned-weight update. Source/math reviewed, locally evaluated by unchanged official sweep and exact repeat."
+  },
+  "command_calibration": {
+    "input_index_zero_based": 48,
+    "guard": "float32(0.3)<=vx_external<=float32(2.0)",
+    "inside_arithmetic": "Clip(float32(vx_external+float32(-0.05)),float32(0.3),float32(2.0))",
+    "outside_arithmetic": "Return original float32 vx_external bits unchanged.",
+    "offset_float32": -0.05000000074505806,
+    "lower_bound_float32": 0.30000001192092896,
+    "upper_bound_float32": 2.0,
+    "other_60_observation_values_unchanged": true,
+    "lateral_and_heading_error_unchanged": true,
+    "normalizer_follows_calibration": true,
+    "actual_previous_action_history": true,
+    "clock_or_seed_conditioning": false,
+    "coast_and_brake_mode_note": "Zero/coast including signed zero and negative/brake values retain original input bits. This preserves the upstream command mode, not a new stopping or hardware-safety guarantee.",
+    "boundary_note": "Discontinuous just above 2: float32(2) maps to about 1.95; its next representable larger input is outside the interval and is unchanged. Positive off-grid values inside the interval are calibrated too."
+  },
+  "local_evaluation": {
+    "date": "2026-10-25",
+    "event": "roller-sprint-2m",
+    "runner_version": 4,
+    "full_official_sweep_races": 30,
+    "mean_finish_s": 3.112,
+    "external_command_m_s": 1.1,
+    "internal_float32_command_m_s": 1.0500000715255737,
+    "seeds_finished": 5,
+    "per_seed_s": [
+      3.1,
+      3.08,
+      3.16,
+      3.08,
+      3.14
+    ],
+    "reference_mean_s": 3.136,
+    "gain_s": 0.024,
+    "sum_timed_ticks_reference": 784,
+    "sum_timed_ticks_candidate": 778,
+    "exact_baseline_arrays": 20,
+    "exact_repeat_arrays": 20,
+    "all_preparation_and_confirmation_races": 45,
+    "all_control_steps_including_settle": 19514,
+    "saved_preflight_ORT_rows": 1818,
+    "candidate_exports": 1,
+    "fresh_seed_robustness_claim": false,
+    "server_submission_verified": false,
+    "independent_saved_audit_sha256": "a82a205e9fd2a9834332df650bb425df48f909ee92fd0873243687269788ce61",
+    "terminal_no_weights_archive_sha256": "bc511c2fa451d23a887e79d13130900983b95ea5f5813325dca9711ea7b42999"
+  },
+  "publication_status": {
+    "state": "published_public_huggingface_not_submitted",
+    "proposed_repo": "djduecaster/dfreq-rollers",
+    "proposed_entry": "dfreq-rollers",
+    "publication_authorized": true
+  },
+  "rights_notice": "Pollen Robotics upstream neural weights are credited; no standalone OSS model license is asserted or invented. This publication is an attributed Hugging Face service-scoped remix under applicable platform terms, not a blanket off-platform redistribution grant. Publication was authorized separately by the user."
+}

policy.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2092a1d8e5b30f2778f227ddf6f26f8eda6487236279f8663bdff71e12805590
+size 794263