Sync policy checkpoints (2026-06-25T14:03:13Z)

ETHRC-humanoidv1-suspended-imu-v0_u4_imu_arm_lg_es500/README.md

@@ -0,0 +1,61 @@
+# Suspended IMU arm-pointing policy — pure-IMU, low-gain, stabilized (`u4_imu_arm_lg_es500`)
+
+A bring-up policy that validates the IMU pipeline end-to-end on ETHRC Humanoid V1 (URDF_4). The
+suspended (fixed-base) robot is tilted continuously and the **arms point down the felt-gravity
+vector**. The policy reads **only the IMU**, so correct pointing proves the IMU observation path is
+wired and consumed correctly — before the IMU is trusted by the walker.
+
+This run supersedes `u4_imu_arm_pure`: it trains at the **updated (lower) arm drive gains** and is
+**early-stopped at the peak** so it does not erode (the long 8000-iter run drifted from ~9.5 down to
+8.2 as entropy crept back up). Exported checkpoint: **`model_175`** (the reward peak).
+
+## Observation (actor) — IMU only
+A real IMU provides exactly these two quantities, and the deployable policy sees nothing else:
+
+| term | meaning | dims | scale |
+|---|---|---|---|
+| `base_ang_vel` | gyro: torso angular velocity in the base frame | 3 | 0.2 |
+| `projected_gravity` | felt-gravity direction in the base frame = `quat_rotate_inverse(base_quat, [0,0,-1])` | 3 | 1.0 |
+
+`history_length = 5` → flattened obs **30**. No joint positions/velocities or last-action, so the
+policy cannot take proprioceptive shortcuts and must derive the pose from the IMU. (The critic uses
+full privileged obs during training only — asymmetric actor-critic.)
+
+## Action — 6 arm joints, position targets
+`shoulder_flexion_{l,r}`, `shoulder_abduction_{l,r}`, `elbow_flexion_{l,r}` (scale 0.5, default
+offset). All other joints are PD-held at default. The action is a **position target**, so the joint
+PD closes the loop and the policy is a pure **IMU → arm-pose** map (sim-agnostic: same IMU → same
+commanded pose in Isaac, MuJoCo, and on hardware). `shoulder_rotation` is intentionally excluded (it
+twists about the arm axis without changing the pointing direction).
+
+## Drive gains (matches the control-repo "test lower gains" update)
+Trained and deployed at the lowered maxon JVPT gains (JVPT = sim gain × 1000):
+
+| group | joints | kp | kd |
+|---|---|---|---|
+| HEJ70 shoulders | `shoulder_flexion`, `shoulder_abduction` | 75 | 1.8 |
+| HEJ50 arms/neck | `shoulder_rotation`, `elbow_flexion`, … | 37.5 | 0.9 |
+
+Legs/waist stay P200 (200/10) and are PD-held; they do not move in this task.
+
+## Training
+Isaac Lab + RSL-RL PPO, 4096 envs, **500 iterations (early-stop)**. The base is driven by a
+continuous, per-env randomized tilt sweep (±45°) so the policy learns to track a *moving* felt-gravity
+vector. Reward: both arm segments (shoulder→elbow and elbow→wrist) aligned with world-down — a dense
+`(cos+1)/2` term plus a tight `exp(-(1-cos)/0.10)` term — averaged over both arms, plus light
+smoothness penalties (`action_rate −0.1`, `joint_acc −2.5e-6`, `energy −2e-5`).
+
+Stability settings (vs the earlier run that eroded): `entropy_coef 0.005`, `desired_kl 0.005` (the
+adaptive LR settles instead of wandering). Result: `point_gravity` peaks **9.63 @ iter 201** and holds
+in a tight 9.4–9.6 band (drop-from-peak 0.24, vs 1.29 before); action std collapses 0.99 → 0.15.
+
+## Files
+- `policy.pt` — TorchScript actor (obs 30 → 6 actions); the obs normalizer is folded in.
+- `policy.onnx` — same network in ONNX.
+- `deploy.yaml` — full deployment config: obs layout (term order/scale/history), per-joint PD gains
+  and limits, action joint order/scale, control rate, and the source checkpoint.
+
+## Use
+Drive it with the production deploy brain (`DeployConfig` + `PolicyController` + `ObservationBuilder`).
+Validate in MuJoCo with `sim2sim_imu_tilt.py` (tilt the base, watch the arms aim down) or feed the
+real Xsens MTi-610 IMU. base_link convention: +X = right, +Y = forward, +Z = up.

ETHRC-humanoidv1-suspended-imu-v0_u4_imu_arm_lg_es500/deploy.yaml

@@ -0,0 +1,261 @@
+experiment_name: ETHRC-humanoidv1-suspended-imu-v0_u4_imu_arm_lg_es500
+task: ETHRC-humanoidv1-suspended-imu-v0
+interface: lo
+sim2sim: true
+msg_type: go
+imu_type: base_link
+lowcmd_topic: rt/lowcmd
+lowstate_topic: rt/lowstate
+policy_path: policy.pt
+num_actions: 6
+action_scale: 0.5
+use_default_offset: true
+control_dt: 0.02
+policy_rate_hz: 50.0
+decimation: 4
+sim_dt: 0.005
+joint_names:
+- shoulder_flexion_l
+- shoulder_flexion_r
+- shoulder_abduction_l
+- shoulder_abduction_r
+- elbow_flexion_l
+- elbow_flexion_r
+joint2motor_idx:
+- 0
+- 1
+- 2
+- 3
+- 4
+- 5
+default_angles:
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+kps:
+- 75.0
+- 75.0
+- 75.0
+- 75.0
+- 37.5
+- 37.5
+kds:
+- 1.7999999523162842
+- 1.7999999523162842
+- 1.7999999523162842
+- 1.7999999523162842
+- 0.8999999761581421
+- 0.8999999761581421
+effort_limits:
+- 80.0
+- 80.0
+- 80.0
+- 80.0
+- 40.0
+- 40.0
+articulation_joint_names:
+- hip_central_rotation
+- waist_upperbody_yaw
+- leg_flexion_l
+- leg_flexion_r
+- neck_yaw
+- shoulder_flexion_l
+- shoulder_flexion_r
+- leg_abduction_l
+- leg_abduction_r
+- neck_pitch
+- shoulder_abduction_l
+- shoulder_abduction_r
+- hip_yaw_l
+- hip_yaw_r
+- shoulder_rotation_l
+- shoulder_rotation_r
+- knee_l
+- knee_r
+- elbow_flexion_l
+- elbow_flexion_r
+- foot_pitch_l
+- foot_pitch_r
+- wrist_rotation_l
+- wrist_rotation_r
+- foot_roll_l
+- foot_roll_r
+articulation_default_angles:
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+- 0.0
+pd_per_joint:
+  hip_central_rotation:
+    kp: 200.0
+    kd: 10.0
+    effort_limit: 128.0
+    velocity_limit: 14.65999984741211
+  waist_upperbody_yaw:
+    kp: 200.0
+    kd: 10.0
+    effort_limit: 128.0
+    velocity_limit: 14.65999984741211
+  leg_flexion_l:
+    kp: 200.0
+    kd: 10.0
+    effort_limit: 128.0
+    velocity_limit: 14.65999984741211
+  leg_flexion_r:
+    kp: 200.0
+    kd: 10.0
+    effort_limit: 128.0
+    velocity_limit: 14.65999984741211
+  leg_abduction_l:
+    kp: 200.0
+    kd: 10.0
+    effort_limit: 128.0
+    velocity_limit: 14.65999984741211
+  leg_abduction_r:
+    kp: 200.0
+    kd: 10.0
+    effort_limit: 128.0
+    velocity_limit: 14.65999984741211
+  hip_yaw_l:
+    kp: 200.0
+    kd: 10.0
+    effort_limit: 128.0
+    velocity_limit: 14.65999984741211
+  hip_yaw_r:
+    kp: 200.0
+    kd: 10.0
+    effort_limit: 128.0
+    velocity_limit: 14.65999984741211
+  knee_l:
+    kp: 200.0
+    kd: 10.0
+    effort_limit: 128.0
+    velocity_limit: 14.65999984741211
+  knee_r:
+    kp: 200.0
+    kd: 10.0
+    effort_limit: 128.0
+    velocity_limit: 14.65999984741211
+  foot_pitch_l:
+    kp: 40.0
+    kd: 1.0
+    effort_limit: 40.0
+    velocity_limit: 15.0
+  foot_pitch_r:
+    kp: 40.0
+    kd: 1.0
+    effort_limit: 40.0
+    velocity_limit: 15.0
+  foot_roll_l:
+    kp: 40.0
+    kd: 1.0
+    effort_limit: 40.0
+    velocity_limit: 15.0
+  foot_roll_r:
+    kp: 40.0
+    kd: 1.0
+    effort_limit: 40.0
+    velocity_limit: 15.0
+  shoulder_flexion_l:
+    kp: 75.0
+    kd: 1.7999999523162842
+    effort_limit: 80.0
+    velocity_limit: 18.0
+  shoulder_flexion_r:
+    kp: 75.0
+    kd: 1.7999999523162842
+    effort_limit: 80.0
+    velocity_limit: 18.0
+  shoulder_abduction_l:
+    kp: 75.0
+    kd: 1.7999999523162842
+    effort_limit: 80.0
+    velocity_limit: 18.0
+  shoulder_abduction_r:
+    kp: 75.0
+    kd: 1.7999999523162842
+    effort_limit: 80.0
+    velocity_limit: 18.0
+  neck_yaw:
+    kp: 37.5
+    kd: 0.8999999761581421
+    effort_limit: 40.0
+    velocity_limit: 20.0
+  neck_pitch:
+    kp: 37.5
+    kd: 0.8999999761581421
+    effort_limit: 40.0
+    velocity_limit: 20.0
+  shoulder_rotation_l:
+    kp: 37.5
+    kd: 0.8999999761581421
+    effort_limit: 40.0
+    velocity_limit: 20.0
+  shoulder_rotation_r:
+    kp: 37.5
+    kd: 0.8999999761581421
+    effort_limit: 40.0
+    velocity_limit: 20.0
+  elbow_flexion_l:
+    kp: 37.5
+    kd: 0.8999999761581421
+    effort_limit: 40.0
+    velocity_limit: 20.0
+  elbow_flexion_r:
+    kp: 37.5
+    kd: 0.8999999761581421
+    effort_limit: 40.0
+    velocity_limit: 20.0
+  wrist_rotation_l:
+    kp: 37.5
+    kd: 0.8999999761581421
+    effort_limit: 40.0
+    velocity_limit: 20.0
+  wrist_rotation_r:
+    kp: 37.5
+    kd: 0.8999999761581421
+    effort_limit: 40.0
+    velocity_limit: 20.0
+obs_history_length: 5
+obs_single_frame_dim: 6
+obs_total_dim: 30
+obs_layout:
+- name: base_ang_vel
+  start: 0
+  end: 15
+  base_dim: 3
+  block_dim: 15
+  scale: 0.2
+- name: projected_gravity
+  start: 15
+  end: 30
+  base_dim: 3
+  block_dim: 15
+  scale: 1.0
+command_ranges: null
+checkpoint: /home/sjohn/Documents/project/rc_humanoid_rl_lab/logs/rsl_rl/humanoid_v1_locomotion/ETHRC-humanoidv1-suspended-imu-v0_u4_imu_arm_lg_es500/model_175.pt

ETHRC-humanoidv1-suspended-imu-v0_u4_imu_arm_lg_es500/policy.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:13c5ec50d2344b8f817a3bddfed723cdef61c7f929003dd2af1e99c2bb5164d5
+size 724625

ETHRC-humanoidv1-suspended-imu-v0_u4_imu_arm_lg_es500/policy.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:562ac44afc4bbd4c350d70273729ad6da1ebfd500b19f75673dac429e076005a
+size 736662

deploy.md

@@ -0,0 +1,55 @@
+# policy_checkpoints — deploy workflow
+
+Deployment checkpoints for the ETHRC humanoid live here, synced with the **private**
+HuggingFace bucket `ETHRC-humanoid/rc-humanoid-locomotion`. Each run is one subfolder
+holding the deploy contract + weights (`deploy.yaml`, `policy.pt`, `policy.onnx[.data]`).
+
+**Nothing in this folder is git-tracked** except the sync scripts and this doc (see
+`.gitignore`) — the weights and configs live on HF and are pulled on demand, so the repo
+never bloats and never drifts from the bucket.
+
+## 0. One-time access (per machine)
+
+You need to be in the `ETHRC-humanoid` HF org. In the training repo, run:
+
+```bash
+cd <rc_humanoid_rl_lab> && ./setup_hf.sh
+```
+
+It tells you to **ask Darius on Discord** for org access, then stores your token
+(`~/.hf_token` + `HF_TOKEN` in your shell profile). On a cluster, just set `HF_TOKEN` as a
+job secret instead. The pull/push scripts here read it automatically.
+
+## 1. Pull the latest checkpoints
+
+```bash
+./pull_from_hf.sh
+```
+
+Downloads every run folder from the bucket into this directory.
+
+## 2. Deploy in sim2sim
+
+The `sim2sim` package drives any pulled checkpoint via `--yaml <RUN_FOLDER>/deploy.yaml`. Use a python
+with torch + mujoco (the `ethrc-deploy` / `env_isaaclab` env). After
+`colcon build --packages-select sim2sim deployment`:
+
+```bash
+# suspended (hang) — bring-up policies (knee/arm/ankle cycle, joint tracking)
+ros2 run sim2sim sim2sim --yaml "$(pwd)/<RUN_FOLDER>/deploy.yaml" --base hang --view
+
+# ground / soft-release (G1 elastic band) — standing / locomotion policies
+ros2 run sim2sim sim2sim --yaml "$(pwd)/<RUN_FOLDER>/deploy.yaml" --base release --view
+```
+
+Without a built workspace, run the module directly:
+`PYTHONPATH=<repo>/ros2_ws/src/sim2sim:<repo>/ros2_ws/src/deployment python -m sim2sim.sim2sim --yaml <…>/deploy.yaml --base hang --view`
+
+## 3. Push new checkpoints
+
+The primary push is from the training repo right after `play_export.py`
+(`<rc_humanoid_rl_lab>/upload_deploy_to_hf.sh`). To push checkpoints curated here:
+
+```bash
+./push_to_hf.sh
+```