Skip to content

armgit / moss

Public Imported from GitHub
0

MOSS: a small, mostly 3D-printed litter-picking rover. Cleaner streets, one maker at a time.

Python
6.6 KB
Raw Blame

Does the MOSS shoulder want a second servo?

MuJoCo studies on the R92 model (../mujoco/model/moss_robot.xml) for issue #2: the shoulder's load at every pose MOSS uses, the pick-and-drop cycle under full dynamics, two STS3215 on one shaft, a spring instead, and tipping. The full write-up with charts and tables is out/report.html (download it and open it in a browser).

Status: simulation evidence, not a bench result. The arm is the model's, scaled to the 768 g weighed in issue #2; servo figures are Feetech's; how stiffly the servo holds its goal, bracket stiffness, encoder offsets and the follower's behaviour are modelled, not measured. The bench steps at the end are what would settle them.

Findings

The litter is MOSS's own test object: the can in ../mujoco/model/scene.xml (66 × 115 mm, 20 g), the one the keep-out table sweeps. The heavier case is an empty 0.33 l glass bottle (250 g), since the detector picks glass by default. Heat is (torque / 0.98 N·m)², the servo's continuous rating (10 kg·cm); stall is 30 kg·cm at 12 V, about 2.65 N·m on the 3S pack.

  1. The litter barely counts. The can adds 0.04 N·m at the pick. Over a whole cycle the shoulder peaks at 50 % of stall and averages 0.15× its rating; the elbow stays under its rating throughout.
  2. The weak moment is holding still at the pick. MOSS picks every object at one pose (pick_low), where the shoulder holds 0.92 N·m, 0.89× its rating, with the can or without it. Over all 645 floor picks of the reach table the median is 1.00× and 40 % are over. The cycle spends about 3 s above the rating there today; a wrist-camera check before the grasp, or a re-grasp, adds time exactly there. Issue #2's test (a can at full reach for a minute) is 2.1× the rating: it can pass without drift while the servo heats.
  3. The cheapest fix is where the arm picks. The arm sits on top of the rover behind its nose, so today's pick is a stretched reach 0.18 m past the tracks' front end (tool at x 0.346 m). The mission drives the object to one pick point, so moving it closer is a config change (MOSS_PICK_XY_M and the pick pose): at x 0.29 m the shoulder runs at 0.46× instead of 0.89×. The pose matters too: at x 0.31 m the reach table's poses range from 0.46× to 1.19×, and its tie-break picks the pose nearest pick_low, not the lightest. Whether the nose camera still sees the object that close is a check on the robot (s6_pick_point).
  4. Whether two servos given the same goal share depends on how stiffly the servo holds its goal. With the model's servo (Menagerie's 998 N·m/rad: full torque about 2 ticks, 0.17°, off its goal), the second servo takes 14 % of the load through a printed bracket when perfectly calibrated, and at 4 ticks of encoder offset the hotter one runs at 2.9×, against 0.89× for one servo; same goal was hotter than one servo in 47 of 84 variants at the pick. The bench note in issue #2 says the real loop is much softer near its goal: under load, a goal a few ticks away does not move the arm, which a 998 N·m/rad servo cannot do. With a servo soft enough that one alone sags 0.5° at the pick (100 N·m/rad), same goal with RoArm's zero at boot beats one servo at any offset up to ±4 ticks (0.74× at worst); at a 1° sag it is 0.53× at worst. One reading on today's arm decides it: one servo at pick_low, Present_Position minus the goal. About 6 ticks or more and same goal will share; 1–2 ticks and it fights.
  5. If the loop is stiff, a follower shares evenly. Servo A holds position; servo B runs open-loop PWM (mode 2) and copies A's Present_Load over the bus. That splits 50/50 at any offset: 0.22× each at the pick. B has to follow with a time constant of about 0.1 s (at 50 Hz, 20-30 % of the gap per update); faster, it rings against A's loop.
  6. For a load that is all arm weight, a spring fits better. A gravity-balance spring of 0.66 N·m peak (zero free length, torque A·sin(θ − φ)) brings the pick pose to 0.08×, with nothing to sync. With a 250 g glass bottle at the pick: one servo 2.0×, spring 0.57×, two servos sharing evenly 0.50× each.
  7. Tipping is not a risk at these weights: 1.09 kg held stretched out to the side before the rover tips on flat ground.

A model note: the XML's shoulder range (±100°) stops short of the drive pose (−104.4°, ticks 860, the calibrated end stop). moss_dimos widens the ranges when it writes its own scene; the XML alone does not.

Bench steps that would settle it

  1. How stiffly one servo holds, on today's arm: at pick_low, the shoulder's Present_Position minus its goal, arriving from above; then step the goal by 2, 4 and 8 ticks and note when the arm moves. About 6 ticks of sag or more: same goal will share. 1–2 ticks: expect a fight, use the follower.
  2. Encoder offset across the travel: both servos on the joint, torque off, swing the arm by hand end to end, log both Present_Position.
  3. Idle fight: torque on, arm in the drive pose, log Present_Load and Present_Current on both for 30 s. Opposite signs mean they fight.
  4. Issue #2's hold with the can, 60 s at pick_low and at full reach, logging Present_Temperature on both: one servo, same goal, follower, spring.
  5. The pick point a few cm closer: does the nose camera still see the object, and does the approach land it there?

Run

The same environment as ../mujoco:

python3 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
make all        # about 25 minutes; writes out/*.json and out/report.html
Study Question
s1_torque_budget Holding torque per joint at every pose MOSS uses (cycle, drive, 645 floor picks, worst stretch), against the servo's continuous and stall torque
s2_cycle The pick-and-drop sequence of config/arm_drop_poses.json under full dynamics, one vs two shoulder servos, payload and a 20 s wait at the pick
s3_two_servos Two STS3215 on one shaft: same goal vs follower, encoder offset × bracket stiffness × play × P; the follower's gain and bus rate; same goal against how stiffly the servo holds its goal
s4_tipping Least payload that tips the rover, and the slope margin
s5_spring A gravity-balance spring instead of (or with) a second servo
s6_pick_point Shoulder load against where the arm picks, straight ahead, from the reach table

moss_arm.py holds the shared model edits (rover pinned, arm scaled to 768 g, payload at the tcp) and the servo figures. Contacts are off in the dynamic studies: they measure what the servos must give, not grasp success.