Arm + Hand record / replay

The teach/replay scripts now capture and replay the Inspire hand alongside the G1 arm. Backward compatible: old recordings (no hand track) replay arm-only.

What changed

g1_teach_v4_stable.py — while recording the arm, it also subscribes to rt/inspire/cmd and saves the commanded finger pose as "h": [12] per frame (0–5 right, 6–11 left). Meta gains "hand": true. On by default; --no-hand disables.
g1_replay_v4_stable.py — auto-detects the h track and publishes it to rt/inspire/cmd every frame, so the hand follows the arm (incl. move-to-start, pause-hold, and home-return).

The hand state read-back is unreliable on the current adapters, so we record the commanded hand pose (rt/inspire/cmd) — i.e. whatever you drive the hand to during teaching (via the dashboard) is what gets captured.

Prerequisites

Env with unitree_sdk2py: conda activate g1_env (also works: teleop, g1ik, …).
Workstation ethernet to the G1 (enp3s0, 192.168.123.x) — same DDS domain as the robot.
inspire_g1 running on the robot (the Docker container) so it receives rt/inspire/cmd and drives the physical hand. Same domain reaches it over the eth link.

Record (teach)

conda activate g1_env
cd ~/Robotics_workspace/yslootahtech/G1/G1_Lootah/Manual_Recorder
python3 g1_teach_v4_stable.py enp3s0 --output wave_hi --seconds 20

First 3 s: arms locked. Then arms go limp — move the arm by hand.
Default (--hand-source keys, terminal focused): o=open · c=close (fist) · b=box · t=thumbs-up — each key drives the hand and is recorded. Reliable.
--hand-source push (push-to-teach): after a 2 s force-baseline calibration, press a finger to move it (via the force sensor, since fingers can't be pushed directly). Two behaviours:
- --push-mode hold (default): press-and-HOLD → the finger closes gradually (harder press = faster, --close-rate); release → holds at that level, so you get any partial close. Press again to open it back.
- --push-mode toggle: a press flips the finger full open↔close.
Keyboard while recording: o/c = open/close both hands · [ ] = open/close right · ; ' = open/close left. Sensitivity: --push-thresh 50 default (auto-raised per finger to its resting noise); lower it (e.g. 30) for a lighter press.
--hand-source cmd: command the hand from the dashboard http://:8088.
--hand-source follow: low-force by-hand — does NOT work on the RH56 (see below).
Arm-only: add --no-hand.

The RH56 fingers ARE back-drivable when RELEASED (corrected 2026-07-30). Sending q=-1 turns the motor OFF (0xFFFF) → the fingers go loose and you can pose them by hand, and the position sensor follows. (An earlier note here wrongly said non-backdrivable — that test used a low grip force, which still COMMANDS/holds a position, not a release.) So the simplest record-by-hand is release + record the actual position — see hand_record.py, which does exactly that. The keys / push / follow teach modes were built before this was understood; keys still works fine, but posing by hand (release) is the natural way.

⚠️ The two hands are on a flaky USB hub that intermittently drops one or both adapters (a dropped hand reads flat 0 / can't be driven). For reliable recording, reseat the USB-RS485 adapters or plug them direct into the Jetson.

Recording fingers by physically moving them (default `--hand-source state`)

rt/inspire/state read-back now works (the frame-header fix — this hand replies 90 EB). In the default state source the recorder releases the fingers (sends -1 each frame so the motors go free) and logs the actual finger degrees, so you can bend the fingers by hand and it captures the real motion. Sources:

--hand-source state (default) — release + record actual positions (move fingers by hand).
--hand-source keys — drive with the keyboard (o/c/b/t) and record that.
--hand-source cmd — record whatever the dashboard commands.

Note: both hands track (open≈1.0 → close≈0), verified 2026-07-29. If a hand reads a flat 0, restart the hand service (./manage.sh restart / re-run the container); if it persists on only one hand, check that hand's RS-485 adapter.

Preview before saving

After the arms return home (and before the save prompt) the teach asks: ▶️ Replay this recording to preview before saving? [y/n]. Answering y replays the just-recorded arm and hand in real time (arms + fingers move), returns to home, and asks again — so you can watch it as many times as you like, then choose save y/n.

Replay (arm + hand together)

conda activate g1_env
python3 g1_replay_v4_stable.py enp3s0 --input wave_hi.jsonl

Detects the hand track automatically and drives the hand with the arm.
Keys: x pause · z resume · q quit. --speed 0.5 for slow-mo.

File format

{"meta":{"hz":60.0,"motors":29,"hand":true,"hand_motors":12}}
{"t":0.0,"q":[..29..],"h":[..12..]}
...

5.1 KiB Raw Permalink Blame History Unescape Escape