fazmain/tendon-arm-runs

Tendon Arm Simulation Runs

Evaluation runs of a simulated, 3D-printable tendon-driven continuum arm (a "Doc Ock" style vertebra spine bent by stepper-motor cables) doing tabletop manipulation in MuJoCo. Each episode is run through Inspect Robots, and every episode is logged in full: multi-camera video, what the policy saw, every action, 100 Hz robot telemetry (all cable tensions, stepper positions, bends, wrist, finger forces, object poses), and a list of detected failure events.

Current contents: 14 runs, 42 episodes. Code: github.com/fazmain/tendon-arm.

The purpose is to find out what this kind of soft-ish, cable-driven arm can and cannot do before building it, and to compare control policies (a scripted baseline with privileged object poses now; Claude models driving the arm from camera images through Inspect Robots next).

The robot

Spine	14 vertebrae at 24 mm pitch in two sections (7 proximal, 7 distal), each joint a lumped 3-axis hinge (±25° bend, ±12° twist) with stiffness and damping
Actuation	8 cables (4 per section, at ±X/±Y), each on a stepper with a 7 mm spool; stall 0.40 N·m, so max tension 57 N; cables pull only
Cable routing	proximal cables at 45 mm radius, distal at 32 mm, rotated 45°; decoupled: proximal bend = r_out·θP, distal = r_in·(θP+θD)
Wrist	roll only, ±150°, 0.25 N·m micro servo
Gripper	parallel jaw, 1 DOF (both fingers coupled), 24 mm travel per finger, 6 N grip, rubber pads with torsional friction
Cameras	front, top (fixed) and L_palm (in the palm, looking along the fingers)
Objects	100 g: a 14 × 2.5 × 2.5 cm bar or 4 cm cubes; green 10 cm target pad; grey 7 × 7 × 4 cm raised block

Units: metres, newtons, seconds; world frame has +z up and the arm base at the origin, with the work area in front of it at y ≈ −0.20 m. Pointing the hand straight down is accurate only in a band 19–22 cm in front of the base (the hand has no wrist pitch; tilt comes from bending the spine).

Tasks

Task	Goal (judged from simulator state, must hold for 1 s)	Time limit
tendon-reach	gripper 6 cm above the pad centre (0.5 s hold)	40 s
tendon-lift	bar ≥ 5 cm above the table, held	90 s
tendon-pick-place	bar resting on the pad, released	120 s
tendon-push	cube slid onto the pad, never lifted	90 s
tendon-rotate	bar turned 90° (±15°) within 5 cm of where it started, released	120 s
tendon-stack	cube resting on the raised block, released	120 s
tendon-sort	blue cube on the pad, orange cube not on it	120 s

Scenes s0, s1, … jitter object positions (±1 cm) and yaw (±15–20°) by seed.

Action space (end-effector mode, 10 Hz): [x, y, z, jaw_yaw_deg, grip] — gripper target in metres, jaw closing direction in degrees, grip 1 = open … 0 = closed. A 100 Hz resolved-rate controller turns this into cable lengths; z ≥ 0.40 parks the arm upright.

Results

policy	batch	task	success	safe (no critical events)	median peak cable load
tendon_scripted	baseline	tendon-lift	3/3	3/3	48%
tendon_scripted	baseline	tendon-pick-place	3/3	3/3	51%
tendon_scripted	baseline	tendon-push	2/3	3/3	82%
tendon_scripted	baseline	tendon-reach	3/3	3/3	49%
tendon_scripted	baseline	tendon-rotate	3/3	3/3	59%
tendon_scripted	baseline	tendon-sort	2/3	3/3	72%
tendon_scripted	baseline	tendon-stack	0/3	2/3	83%
tendon_scripted	wristfix	tendon-lift	3/3	3/3	48%
tendon_scripted	wristfix	tendon-pick-place	3/3	3/3	49%
tendon_scripted	wristfix	tendon-push	3/3	2/3	90%
tendon_scripted	wristfix	tendon-reach	3/3	3/3	49%
tendon_scripted	wristfix	tendon-rotate	3/3	3/3	56%
tendon_scripted	wristfix	tendon-sort	3/3	3/3	60%
tendon_scripted	wristfix	tendon-stack	0/3	3/3	66%

tendon_scripted is a waypoint controller that reads the true object pose from the simulator, so it is an upper-bound-ish baseline, not a learned policy. Batches: baseline is the first full pass; wristfix is after fixing a wrist-roll controller bug (the wrist integrated its own command, wound up, and thrashed between its limits while the arm moved). Stack still fails: about 1 cm of hand placement error is too much to set a 4 cm cube on a 7 cm block.

Layout

episodes.csv                      one row per episode (the default dataset config)
runs/<timestamp>_<task>_<policy>[_<batch>]/
  report.md                       scores, per-episode table, critical events, file index
  command.txt, console.log        exact command and full console output
  transcripts.txt                 agent conversation (model policies; trivial for scripted)
  wire.txt                        every model API request/response (model policies)
  inspect/<task>_<id>.json        Inspect Robots eval log
  inspect/report.html             Inspect Robots HTML viewer, self-contained
  episodes/<NNN>_<scene>/
    episode.mp4                   30 fps, all cameras tiled, caption with time, cable load, active failures
    policy_view.mp4               exactly the images the policy received, one frame per action
    actions.csv                   every action, with success flag and new/active failure events
    telemetry.csv                 100 Hz robot and object state (columns below)
    events.jsonl                  one line per failure event
    summary.json                  outcome, durations, peaks, event counts, final object poses
    scene.json                    seed, object start poses, targets, full arm parameters
    model.xml                     the exact MuJoCo model for this episode (reproducible)

Rerun recordings (.rrd) are not uploaded; they duplicate the CSVs and videos.

episodes.csv

run, timestamp, task, policy, tag, episode, scene_id, success, end_reason, sim_seconds, control_steps, critical_events, peak_cable_load_pct, event_counts (JSON), video (path). end_reason is success, fault, object_lost, reset/closed (time or step limit), or the policy's done/give_up.

telemetry.csv (100 Hz, arm prefix L_)

Column	Meaning
t	simulation time, s
L_T_{P,D}{+X,-X,+Y,-Y}	tension in each of the 8 cables, N (stall = 57.1 N)
L_motor_mm_*, L_cmd_mm_*	stepper cable position actually reached vs commanded, mm
L_theta_cmd_{Px,Py,Dx,Dy}, L_theta_meas_*	section bend commanded vs measured, rad (total over the section)
L_wrist_cmd_deg, L_wrist_deg, L_wrist_torque	wrist roll setpoint, angle, torque (N·m)
L_grip_cmd, L_finger_gap_mm	grip command (fraction open), gap between pads
L_finger_a_N, L_finger_b_N, L_grip_servo_N	contact force on each finger, gripper servo force
L_grip_{x,y,z}, L_target_{x,y,z}	gripper centre and the controller's current target, m
L_hand_tilt_deg	angle of the hand axis from vertical up (180 = pointing straight down)
L_max_joint_deg, L_twist_deg	largest single-joint bend and twist
L_spine_load_N	compression along the spine from all cable tensions
<object>_{x,y,z,qw,qx,qy,qz,vx,vy,vz}	pose and linear velocity of each object
n_contacts	active contacts in the simulation

events.jsonl

{"type": "cable_stall", "severity": "critical", "key": "cable_stall:L:D+X", "t_start": 7.31,
 "t_end": 9.71, "duration": 2.4, "peak": 57.03, "arm": "L", "cable": "D+X"}

A condition becomes an event after persisting 50 ms; gaps under 0.1 s merge into one event.

Event	Severity	Trigger
cable_stall	critical	a cable at ≥ 98% of stall tension (a real stepper would skip steps)
cable_overload	warning	any cable ≥ 85% of stall
stepper_lag	warning	commanded cable length > 4 mm ahead of the stepper
joint_limit, twist_limit	warning	a vertebra joint within 0.5° of its bend/twist limit
wrist_stalled	warning	wrist at ≥ 98% torque and not moving
spine_overload	warning	spine compression > 150 N
target_unreachable	warning	gripper > 3 cm from its target for > 2 s
arm_arm_contact	critical	two arms touching (two-arm layout)
arm_table_contact	warning	spine or hand (not finger pads) touching table or block
finger_table_press	warning	fingers pressing the table with > 5 N
object_bumped	info	arm touched an object with something other than the finger pads
object_tipped	warning	object tilted > 30°
object_dropped	critical	held ≥ 0.3 s, then lost contact while still commanded closed, off the table
object_out_of_workspace	critical	object > 0.6 m from the base
sim_unstable	critical	NaN or simulator warning

The tendon_safety score is 1 when an episode has no critical events.

Usage

import pandas as pd
from huggingface_hub import hf_hub_download

eps = pd.read_csv("hf://datasets/fazmain/tendon-arm-runs/episodes.csv")
row = eps[eps.task == "tendon-pick-place"].iloc[0]
tel = pd.read_csv(f"hf://datasets/fazmain/tendon-arm-runs/{row.video.replace('episode.mp4', 'telemetry.csv')}")
tel.filter(like="L_T_").max()        # peak tension per cable
video = hf_hub_download("fazmain/tendon-arm-runs", row.video, repo_type="dataset")

Or the whole thing: hf download fazmain/tendon-arm-runs --repo-type dataset --local-dir tendon-arm-runs.

Caveats

• Simulation only. The model has not been checked against a physical build; cable friction, stepper dynamics and print compliance are approximations.
• The spine joints are lumped rigid hinges with spring/damper, not a continuum model.
• Scripted-policy runs use privileged object poses; model policies see only the camera images and proprioception listed in each run's Inspect log.