sudarshanharithas commited on
Commit
d375952
·
verified ·
1 Parent(s): c13c941

Publish clean BRACE source and procedural-fixture validation

Browse files
README.md CHANGED
@@ -21,7 +21,6 @@ tags:
21
  [![Project page](https://img.shields.io/badge/Project-Page-2563eb)](https://multiplylabs.github.io/brace/)
22
  [![Interactive demo](https://img.shields.io/badge/Try-Interactive_Demo-2563eb)](https://multiplylabs.github.io/robogym-online/)
23
  [![Video](https://img.shields.io/badge/YouTube-Video-ff0000)](https://www.youtube.com/watch?v=jASGzq4K5m0)
24
- [![Internal GitHub repository](https://img.shields.io/badge/GitHub-Internal_Testing-24292f)](https://github.com/multiplylabs/brace-internal)
25
 
26
  <img src="media/teaser.png" alt="BRACE force exertion, compensation, terrain adaptation and teleoperation on the Unitree G1" width="100%" />
27
 
@@ -29,9 +28,8 @@ tags:
29
 
30
  BRACE adapts a humanoid's whole-body motion reference to commanded hand forces, moments and terrain. Two privileged teachers transfer exertion and compensation behavior into one student. The student runs from proprioception and reference commands without terrain maps or measured hand wrenches.
31
 
32
- This release contains **all three trained checkpoints**, their frozen configurations, SafeTensors weights, the student ONNX export, and a runnable source snapshot. The [paper](https://arxiv.org/abs/2610.07052), [project page](https://multiplylabs.github.io/brace/) and [companion GitHub repository](https://github.com/multiplylabs/brace) describe the method and demonstrations. The source snapshot is included so model use does not depend on the timing of the GitHub code release.
33
 
34
- Internal development and testing use **[multiplylabs/brace-internal](https://github.com/multiplylabs/brace-internal)**, a private GitHub repository requiring team access. The model downloads and bundled source here are public.
35
 
36
  | Model | Files | Observations and role |
37
  |---|---|---|
@@ -64,12 +62,12 @@ pip install -e '.[sim,export,deploy]'
64
  pip install --no-deps -e vendor/ProtoMotions
65
  ```
66
 
67
- For headless Ubuntu, install `libosmesa6`, `libgl1`, and `libegl1` and set `MUJOCO_GL=osmesa` and `PYOPENGL_PLATFORM=osmesa` before simulation imports. Supply a compatible packed G1 MotionLib; motion data is distributed separately. From the extracted `brace/` directory:
68
 
69
  ```bash
70
  python -m brace_repro sim2sim --stage student --mode exert \
71
  --checkpoint ../brace-models/artifacts/student/last.ckpt \
72
- --motion /path/to/g1_motion_library.pt \
73
  --force 3 0 0 --torque 0.1 0 0 \
74
  --steps 100 --output runs/student-exert
75
  ```
@@ -93,24 +91,24 @@ Teachers require privileged observations and run in simulation. Use the correspo
93
  ```bash
94
  python -m brace_repro sim2sim --stage exert --mode exert \
95
  --checkpoint ../brace-models/artifacts/teachers/exert/last.ckpt \
96
- --motion /path/to/g1_motion_library.pt \
97
  --force 3 0 0 --torque 0.1 0 0 \
98
  --steps 100 --output runs/teacher-exert
99
 
100
  python -m brace_repro sim2sim --stage comp --mode comp \
101
  --checkpoint ../brace-models/artifacts/teachers/comp/last.ckpt \
102
- --motion /path/to/g1_motion_library.pt \
103
  --force 3 0 0 --torque 0.1 0 0 \
104
  --steps 100 --output runs/teacher-comp
105
  ```
106
 
107
- Teacher training scripts are `scripts/train_exert_teacher.sh` and `scripts/train_comp_teacher.sh`. Student distillation uses `scripts/distill_student.sh` with explicit `--teacher-exert` and `--teacher-comp` paths to these downloaded checkpoints. For fresh training or fine-tuning, follow `docs/training.md` in the source snapshot; training needs Isaac Sim 5.1.0, IsaacLab 2.3.0, and an NVIDIA GPU. Pretrained models do not require Anyscale or a private Seahorse checkout.
108
 
109
  ## Data, validation and limitations
110
 
111
- Training used a packed G1 locomotion/upper-body manipulation motion library with 58,028 clips, terrain variation, commanded hand wrenches and external disturbances. The full motion library is not included in this model release. Replacing it with another library changes the training experiment. The original bundles retain inference configs, weights and training counters, rather than complete original training logs or simulator/RNG state; exact historical training replay is not established.
112
 
113
- Release checks passed 52 tests, 19 functional pipeline cases including short GPU training in all three stages, and six deployment simulation cases totaling 650 steps. The source snapshot also passed source and clean CPU simulation/deployment CI. Checkpoint tensors are unchanged; the SafeTensors exports were checked tensor by tensor. The packaged student passed a further 20-step MuJoCo/ONNX parity check and deployment preflight. These short checks do not establish held-out performance, complete retraining or physical robot readiness. See [validation.json](validation.json) for the release record.
114
 
115
  The robot is the 29-DoF Unitree G1. Teachers use privileged simulator inputs and are not hardware policies. Student deployment requires compatible robot assets, PD gains, motion references, online reference adaptation and state/history conventions. PyTorch checkpoint/config files use pickle; use the included BRACE loader for these trusted release files. SafeTensors provides a pickle-free copy of every model tensor.
116
 
 
21
  [![Project page](https://img.shields.io/badge/Project-Page-2563eb)](https://multiplylabs.github.io/brace/)
22
  [![Interactive demo](https://img.shields.io/badge/Try-Interactive_Demo-2563eb)](https://multiplylabs.github.io/robogym-online/)
23
  [![Video](https://img.shields.io/badge/YouTube-Video-ff0000)](https://www.youtube.com/watch?v=jASGzq4K5m0)
 
24
 
25
  <img src="media/teaser.png" alt="BRACE force exertion, compensation, terrain adaptation and teleoperation on the Unitree G1" width="100%" />
26
 
 
28
 
29
  BRACE adapts a humanoid's whole-body motion reference to commanded hand forces, moments and terrain. Two privileged teachers transfer exertion and compensation behavior into one student. The student runs from proprioception and reference commands without terrain maps or measured hand wrenches.
30
 
31
+ This release contains **all three trained checkpoints**, their frozen configurations, SafeTensors weights, the student ONNX export, and a runnable source snapshot. The [paper](https://arxiv.org/abs/2610.07052), [project page](https://multiplylabs.github.io/brace/) and [companion GitHub repository](https://github.com/multiplylabs/brace) describe the method and demonstrations. The included source snapshot provides the same BRACE runtime and training entrypoints.
32
 
 
33
 
34
  | Model | Files | Observations and role |
35
  |---|---|---|
 
62
  pip install --no-deps -e vendor/ProtoMotions
63
  ```
64
 
65
+ For headless Ubuntu, install `libosmesa6`, `libgl1`, and `libegl1` and set `MUJOCO_GL=osmesa` and `PYOPENGL_PLATFORM=osmesa` before simulation imports. For an execution check, use the included procedural standing fixture. The original recorded motion library is not distributed. From the extracted `brace/` directory:
66
 
67
  ```bash
68
  python -m brace_repro sim2sim --stage student --mode exert \
69
  --checkpoint ../brace-models/artifacts/student/last.ckpt \
70
+ --motion ../brace-models/data/motions/smoke.pt \
71
  --force 3 0 0 --torque 0.1 0 0 \
72
  --steps 100 --output runs/student-exert
73
  ```
 
91
  ```bash
92
  python -m brace_repro sim2sim --stage exert --mode exert \
93
  --checkpoint ../brace-models/artifacts/teachers/exert/last.ckpt \
94
+ --motion ../brace-models/data/motions/smoke.pt \
95
  --force 3 0 0 --torque 0.1 0 0 \
96
  --steps 100 --output runs/teacher-exert
97
 
98
  python -m brace_repro sim2sim --stage comp --mode comp \
99
  --checkpoint ../brace-models/artifacts/teachers/comp/last.ckpt \
100
+ --motion ../brace-models/data/motions/smoke.pt \
101
  --force 3 0 0 --torque 0.1 0 0 \
102
  --steps 100 --output runs/teacher-comp
103
  ```
104
 
105
+ Teacher training scripts are `scripts/train_exert_teacher.sh` and `scripts/train_comp_teacher.sh`. Student distillation uses `scripts/distill_student.sh` with explicit `--teacher-exert` and `--teacher-comp` paths to these downloaded checkpoints. For fresh training or fine-tuning, follow `docs/training.md` in the source snapshot; training needs Isaac Sim 5.1.0, IsaacLab 2.3.0, and an NVIDIA GPU. Pretrained models do not require job runtime or another research checkout.
106
 
107
  ## Data, validation and limitations
108
 
109
+ Training used a packed G1 locomotion/upper-body manipulation motion library with 58,028 clips, terrain variation, commanded hand wrenches and external disturbances. The full motion library is not included in this model release, and its redistribution rights remain unverified. The included `data/motions/smoke.pt` is a procedurally generated static standing reference, not a training-data subset. Its generator uses only the robot definition and standing joint targets and reads no recorded motions. It is provided under Apache-2.0 for execution checks, with provenance in `data/motions/smoke.provenance.json`. It is not a performance benchmark or a replacement training library. Replacing the original training data changes the training experiment. The original bundles retain inference configs, weights and training counters, rather than complete original training logs or simulator/RNG state; exact historical training replay is not established.
110
 
111
+ The cleaned public checkout passed 50 tests, 15 CPU pipeline cases with the procedural fixture, and six deployment simulation cases totaling 650 steps. Earlier runtime checks also included short GPU training in all three stages; those earlier checks used local training data that is not distributed. Checkpoint tensors are unchanged; the SafeTensors exports were checked tensor by tensor. The packaged student passed deployment preflight and a 100-step MuJoCo/ONNX combined-wrench check with the procedural fixture. These short checks do not establish held-out performance, complete retraining or physical robot readiness. See [validation.json](validation.json) for the release record.
112
 
113
  The robot is the 29-DoF Unitree G1. Teachers use privileged simulator inputs and are not hardware policies. Student deployment requires compatible robot assets, PD gains, motion references, online reference adaptation and state/history conventions. PyTorch checkpoint/config files use pickle; use the included BRACE loader for these trusted release files. SafeTensors provides a pickle-free copy of every model tensor.
114
 
artifact_manifest.json CHANGED
@@ -20,8 +20,8 @@
20
  },
21
  {
22
  "path": "README.md",
23
- "bytes": 8158,
24
- "sha256": "6f386925f937a857c3010c6f2a2f317c4da78c4e6738b53e28e8bf350ed1fb88"
25
  },
26
  {
27
  "path": "artifacts/student/README.md",
@@ -120,13 +120,13 @@
120
  },
121
  {
122
  "path": "source/brace-source.tar.gz",
123
- "bytes": 39623826,
124
- "sha256": "b4eeba5e23632e842afa1d5d6483387da0579657ef42336fe1fcd686e3782f62"
125
  },
126
  {
127
  "path": "validation.json",
128
- "bytes": 1757,
129
- "sha256": "17dcb3545e9a719810f30983f95b3d2e6d8223a245908f70059367b8627f0ba6"
130
  },
131
  {
132
  "path": "data/motions/smoke.pt",
 
20
  },
21
  {
22
  "path": "README.md",
23
+ "bytes": 8385,
24
+ "sha256": "00db3861c57ae5a49c6667530b2aff56ab52ec22aecb811e1eaa1454db7d7065"
25
  },
26
  {
27
  "path": "artifacts/student/README.md",
 
120
  },
121
  {
122
  "path": "source/brace-source.tar.gz",
123
+ "bytes": 39631536,
124
+ "sha256": "d3af99a345aced8ecb1010e8484238d7f2a498a9975deae1a05b80e635460606"
125
  },
126
  {
127
  "path": "validation.json",
128
+ "bytes": 9695,
129
+ "sha256": "30e3f4040274e9dcc80b73bc6e552006a7e1b03236b5cf8cfb84e8685c5d3d87"
130
  },
131
  {
132
  "path": "data/motions/smoke.pt",
source/brace-source.tar.gz CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b4eeba5e23632e842afa1d5d6483387da0579657ef42336fe1fcd686e3782f62
3
- size 39623826
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d3af99a345aced8ecb1010e8484238d7f2a498a9975deae1a05b80e635460606
3
+ size 39631536
validation.json CHANGED
@@ -1,56 +1,420 @@
1
  {
2
- "tests_passed": 45,
3
- "pipeline_cases_passed": 19,
4
- "checkpoint_checks": {
5
- "exert": {
6
- "checkpoint_state_preserved": true,
7
- "safetensors_identical": true,
8
- "strict_model_load": true,
9
- "epoch": 8490,
10
- "step_count": 4451205120
11
- },
12
- "comp": {
13
- "checkpoint_state_preserved": true,
14
- "safetensors_identical": true,
15
- "strict_model_load": true,
16
- "epoch": 5700,
17
- "step_count": 2988441600
18
- },
19
- "student": {
20
- "checkpoint_state_preserved": true,
21
- "safetensors_identical": true,
22
- "strict_model_load": true,
23
- "epoch": 900,
24
- "step_count": 471859200
25
- }
26
  },
27
- "full_retraining_completed": false,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
28
  "hardware_execution_validated": false,
29
- "source_revision": "f94bb92cc6a136381e10859eb3e880001581e39b",
30
- "packaging_tests_passed": 52,
31
- "packaged_student_deployment_preflight": true,
32
- "packaged_student_sim2sim": {
33
- "kind": "functional_smoke",
34
- "steps": 20,
35
- "num_envs": 1,
36
- "seed": 0,
37
- "finite_actions_and_rewards": true,
38
- "termination_count": 0,
39
- "reset_count": 0,
40
- "mean_metrics": {
41
- "reward": 8.321679377555848,
42
- "hand_force/force_cmd_norm": 3.0,
43
- "hand_force/force_meas_norm": 3.168040084838867,
44
- "hand_force/torque_cmd_norm": 0.24494898319244385,
45
- "hand_force/torque_meas_norm": 0.6538496252149344,
46
- "hand_disturbance/force_ext_w_norm": 0.0,
47
- "hand_disturbance/torque_ext_w_norm": 0.0,
48
- "task_mode/force_cmd_eff_norm": 3.0,
49
- "task_mode/torque_cmd_eff_norm": 0.24494898989796637,
50
- "task_mode/wrench_priv_norm": 3.168040084838867,
51
- "task_mode/torque_priv_norm": 0.6538496438413859
52
- },
53
- "policy_runtime": "onnxruntime",
54
- "live_action_max_abs_difference": 9.213387966156006e-05
55
- }
56
  }
 
1
  {
2
+ "source": "independent public checkout",
3
+ "tests_passed": 50,
4
+ "public_runtime_artifacts": 13,
5
+ "public_url_audit_passed": true,
6
+ "full_training_distribution_available": false,
7
+ "fixture": {
8
+ "kind": "procedural_static_standing",
9
+ "recorded_motion_data_used": false,
10
+ "license": "Apache-2.0"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
11
  },
12
+ "cpu_pipeline": {
13
+ "kind": "functional_pipeline_check",
14
+ "cases": [
15
+ {
16
+ "command": [
17
+ "doctor",
18
+ "--deep"
19
+ ],
20
+ "exit_code": 0
21
+ },
22
+ {
23
+ "command": [
24
+ "sim2sim",
25
+ "--steps",
26
+ "100",
27
+ "--output",
28
+ "<repo>/runs/public-release-cpu/mujoco-zero"
29
+ ],
30
+ "exit_code": 0
31
+ },
32
+ {
33
+ "command": [
34
+ "sim2sim",
35
+ "--mode",
36
+ "exert",
37
+ "--force",
38
+ "3",
39
+ "0",
40
+ "0",
41
+ "--steps",
42
+ "100",
43
+ "--output",
44
+ "<repo>/runs/public-release-cpu/mujoco-exert"
45
+ ],
46
+ "exit_code": 0
47
+ },
48
+ {
49
+ "command": [
50
+ "sim2sim",
51
+ "--mode",
52
+ "comp",
53
+ "--force",
54
+ "6",
55
+ "0",
56
+ "0",
57
+ "--steps",
58
+ "100",
59
+ "--output",
60
+ "<repo>/runs/public-release-cpu/mujoco-comp"
61
+ ],
62
+ "exit_code": 0
63
+ },
64
+ {
65
+ "command": [
66
+ "export",
67
+ "--output",
68
+ "<repo>/runs/public-release-cpu/export"
69
+ ],
70
+ "exit_code": 0
71
+ },
72
+ {
73
+ "command": [
74
+ "sim2sim",
75
+ "--onnx",
76
+ "<repo>/runs/public-release-cpu/export/unified_pipeline.onnx",
77
+ "--steps",
78
+ "100",
79
+ "--output",
80
+ "<repo>/runs/public-release-cpu/mujoco-onnx"
81
+ ],
82
+ "exit_code": 0
83
+ },
84
+ {
85
+ "command": [
86
+ "evaluate",
87
+ "--simulator",
88
+ "mujoco",
89
+ "--mode",
90
+ "exert",
91
+ "--force",
92
+ "3",
93
+ "0",
94
+ "0",
95
+ "--override",
96
+ "agent.evaluator.max_eval_steps=100",
97
+ "--output",
98
+ "<repo>/runs/public-release-cpu/perclip"
99
+ ],
100
+ "exit_code": 0
101
+ },
102
+ {
103
+ "command": [
104
+ "sim2sim",
105
+ "--stage",
106
+ "exert",
107
+ "--mode",
108
+ "exert",
109
+ "--torque",
110
+ ".1",
111
+ ".2",
112
+ "-.1",
113
+ "--steps",
114
+ "100",
115
+ "--output",
116
+ "<repo>/runs/public-release-cpu/wrench-exert-exert-torque"
117
+ ],
118
+ "exit_code": 0
119
+ },
120
+ {
121
+ "command": [
122
+ "sim2sim",
123
+ "--stage",
124
+ "exert",
125
+ "--mode",
126
+ "exert",
127
+ "--torque",
128
+ ".1",
129
+ ".2",
130
+ "-.1",
131
+ "--steps",
132
+ "100",
133
+ "--output",
134
+ "<repo>/runs/public-release-cpu/wrench-exert-exert-combined",
135
+ "--force",
136
+ "3",
137
+ "1",
138
+ "-.5"
139
+ ],
140
+ "exit_code": 0
141
+ },
142
+ {
143
+ "command": [
144
+ "sim2sim",
145
+ "--stage",
146
+ "comp",
147
+ "--mode",
148
+ "comp",
149
+ "--torque",
150
+ ".1",
151
+ ".2",
152
+ "-.1",
153
+ "--steps",
154
+ "100",
155
+ "--output",
156
+ "<repo>/runs/public-release-cpu/wrench-comp-comp-torque"
157
+ ],
158
+ "exit_code": 0
159
+ },
160
+ {
161
+ "command": [
162
+ "sim2sim",
163
+ "--stage",
164
+ "comp",
165
+ "--mode",
166
+ "comp",
167
+ "--torque",
168
+ ".1",
169
+ ".2",
170
+ "-.1",
171
+ "--steps",
172
+ "100",
173
+ "--output",
174
+ "<repo>/runs/public-release-cpu/wrench-comp-comp-combined",
175
+ "--force",
176
+ "3",
177
+ "1",
178
+ "-.5"
179
+ ],
180
+ "exit_code": 0
181
+ },
182
+ {
183
+ "command": [
184
+ "sim2sim",
185
+ "--stage",
186
+ "student",
187
+ "--mode",
188
+ "exert",
189
+ "--torque",
190
+ ".1",
191
+ ".2",
192
+ "-.1",
193
+ "--steps",
194
+ "100",
195
+ "--output",
196
+ "<repo>/runs/public-release-cpu/wrench-student-exert-torque",
197
+ "--onnx",
198
+ "<repo>/runs/public-release-cpu/export/unified_pipeline.onnx"
199
+ ],
200
+ "exit_code": 0
201
+ },
202
+ {
203
+ "command": [
204
+ "sim2sim",
205
+ "--stage",
206
+ "student",
207
+ "--mode",
208
+ "exert",
209
+ "--torque",
210
+ ".1",
211
+ ".2",
212
+ "-.1",
213
+ "--steps",
214
+ "100",
215
+ "--output",
216
+ "<repo>/runs/public-release-cpu/wrench-student-exert-combined",
217
+ "--force",
218
+ "3",
219
+ "1",
220
+ "-.5",
221
+ "--onnx",
222
+ "<repo>/runs/public-release-cpu/export/unified_pipeline.onnx"
223
+ ],
224
+ "exit_code": 0
225
+ },
226
+ {
227
+ "command": [
228
+ "sim2sim",
229
+ "--stage",
230
+ "student",
231
+ "--mode",
232
+ "comp",
233
+ "--torque",
234
+ ".1",
235
+ ".2",
236
+ "-.1",
237
+ "--steps",
238
+ "100",
239
+ "--output",
240
+ "<repo>/runs/public-release-cpu/wrench-student-comp-torque",
241
+ "--onnx",
242
+ "<repo>/runs/public-release-cpu/export/unified_pipeline.onnx"
243
+ ],
244
+ "exit_code": 0
245
+ },
246
+ {
247
+ "command": [
248
+ "sim2sim",
249
+ "--stage",
250
+ "student",
251
+ "--mode",
252
+ "comp",
253
+ "--torque",
254
+ ".1",
255
+ ".2",
256
+ "-.1",
257
+ "--steps",
258
+ "100",
259
+ "--output",
260
+ "<repo>/runs/public-release-cpu/wrench-student-comp-combined",
261
+ "--force",
262
+ "3",
263
+ "1",
264
+ "-.5",
265
+ "--onnx",
266
+ "<repo>/runs/public-release-cpu/export/unified_pipeline.onnx"
267
+ ],
268
+ "exit_code": 0
269
+ }
270
+ ],
271
+ "paper_performance_reproduced": false
272
+ },
273
+ "deployment": {
274
+ "hardware_opened": false,
275
+ "cases": [
276
+ {
277
+ "mode": "exert",
278
+ "force_xyz": [
279
+ 0.0,
280
+ 0.0,
281
+ 0.0
282
+ ],
283
+ "torque_xyz": [
284
+ 0.0,
285
+ 0.0,
286
+ 0.0
287
+ ],
288
+ "reference": "clip",
289
+ "steps": 100,
290
+ "finite_targets_and_state": true,
291
+ "streamed": false,
292
+ "force_gate_seen": true,
293
+ "brace_goal_seen": false,
294
+ "mean_step_ms": 2.8709147591143847,
295
+ "student_torque_command_max_norm": 0.0,
296
+ "root_height_end_m": 0.7717454025206038
297
+ },
298
+ {
299
+ "mode": "exert",
300
+ "force_xyz": [
301
+ 3.0,
302
+ 0.0,
303
+ 0.0
304
+ ],
305
+ "torque_xyz": [
306
+ 0.1,
307
+ 0.2,
308
+ -0.1
309
+ ],
310
+ "reference": "clip",
311
+ "steps": 100,
312
+ "finite_targets_and_state": true,
313
+ "streamed": false,
314
+ "force_gate_seen": true,
315
+ "brace_goal_seen": true,
316
+ "mean_step_ms": 27.275491929613054,
317
+ "student_torque_command_max_norm": 0.34641018509864807,
318
+ "root_height_end_m": 0.771555535822297
319
+ },
320
+ {
321
+ "mode": "exert",
322
+ "force_xyz": [
323
+ 0.0,
324
+ 0.0,
325
+ 0.0
326
+ ],
327
+ "torque_xyz": [
328
+ 0.1,
329
+ 0.2,
330
+ -0.1
331
+ ],
332
+ "reference": "clip",
333
+ "steps": 100,
334
+ "finite_targets_and_state": true,
335
+ "streamed": false,
336
+ "force_gate_seen": true,
337
+ "brace_goal_seen": true,
338
+ "mean_step_ms": 27.242134183179587,
339
+ "student_torque_command_max_norm": 0.34641018509864807,
340
+ "root_height_end_m": 0.7717985918682772
341
+ },
342
+ {
343
+ "mode": "comp",
344
+ "force_xyz": [
345
+ 0.0,
346
+ 0.0,
347
+ 0.0
348
+ ],
349
+ "torque_xyz": [
350
+ 0.0,
351
+ 0.0,
352
+ 0.0
353
+ ],
354
+ "reference": "clip",
355
+ "steps": 100,
356
+ "finite_targets_and_state": true,
357
+ "streamed": false,
358
+ "force_gate_seen": true,
359
+ "brace_goal_seen": false,
360
+ "mean_step_ms": 2.6857469975948334,
361
+ "student_torque_command_max_norm": 0.0,
362
+ "root_height_end_m": 0.7625857493233349
363
+ },
364
+ {
365
+ "mode": "comp",
366
+ "force_xyz": [
367
+ 3.0,
368
+ 0.0,
369
+ 0.0
370
+ ],
371
+ "torque_xyz": [
372
+ 0.1,
373
+ 0.2,
374
+ -0.1
375
+ ],
376
+ "reference": "clip",
377
+ "steps": 100,
378
+ "finite_targets_and_state": true,
379
+ "streamed": false,
380
+ "force_gate_seen": true,
381
+ "brace_goal_seen": false,
382
+ "mean_step_ms": 2.5799738988280296,
383
+ "student_torque_command_max_norm": 0.0,
384
+ "root_height_end_m": 0.7623849673709036
385
+ },
386
+ {
387
+ "mode": "exert",
388
+ "force_xyz": [
389
+ 3.0,
390
+ 0.0,
391
+ 0.0
392
+ ],
393
+ "torque_xyz": [
394
+ 0.1,
395
+ 0.2,
396
+ -0.1
397
+ ],
398
+ "reference": "loopback_vr",
399
+ "steps": 150,
400
+ "finite_targets_and_state": true,
401
+ "streamed": true,
402
+ "force_gate_seen": true,
403
+ "brace_goal_seen": true,
404
+ "mean_step_ms": 20.770735990566514,
405
+ "student_torque_command_max_norm": 0.34641018509864807,
406
+ "root_height_end_m": 0.771784665720328
407
+ }
408
+ ]
409
+ },
410
+ "all_upstream_commits_publicly_verified": true,
411
+ "history_policy": "One code release commit on existing public history; no private history or tags imported.",
412
  "hardware_execution_validated": false,
413
+ "full_retraining_completed": false,
414
+ "paper_performance_reproduced": false,
415
+ "source_archive_relocation_verified": true,
416
+ "publishing_credential_in_source": false,
417
+ "checkpoint_state_preserved": true,
418
+ "safetensors_identical": true,
419
+ "source_snapshot_sha256": "d3af99a345aced8ecb1010e8484238d7f2a498a9975deae1a05b80e635460606"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
420
  }