session string | arm string | round int64 | prompt_system string | prompt_user string | prompt_est_tokens int64 | analysis string | hypothesis string | change_type string | combined_justification string | knobs dict | edits list | new_files list | expected dict | reasoning string | diff string | validation_error null | commit string | branch string | ckpt_repo string | checkpoint_scores list | trained bool | best_checkpoint int64 | score_s float64 | seeds_finished int64 | command_vx float64 | report string | warnings list | expected_score_s float64 | prediction_error_s float64 | tokens_in int64 | tokens_out int64 | reasoning_tokens int64 | beam_attempts int64 | beam_model string | beam_wall_s float64 | finish_reason string | job_minutes float64 | platform string | arena_version string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
A | A | 1 | null | null | null | null | null | null | null | null | null | null | null | null | null | null | 4b675ebf4 | null | munish0838/mjlab-sprint2m-microduck-A-r1 | [
{
"checkpoint": 250,
"sweep": [
{
"command_vx": 0.2,
"status": "fall",
"time_s": null,
"progress_m": 1.56
},
{
"command_vx": 0.25,
"status": "finish",
"time_s": 12.5,
"progress_m": null
},
{
"command_vx": 0... | true | 750 | 5.438 | 5 | 0.5 | ROUND 1 — run A-r1, commit 4b675ebf4
hypothesis: —
change: no change (pristine recipe)
ARENA (local, Arena rules, platform: arm64 Darwin, arena 0.6.0)
checkpoint score finished best vx
250 12.565 5/5 0.25
500 7.509 5/5 0.50
750 5.438 5/5 0.50
999 5.440... | [] | null | null | null | null | null | null | null | null | null | 29 | arm64 Darwin | 0.6.0 |
A2 | A2 | 1 | null | null | null | null | null | null | null | null | null | null | null | null | null | null | 4b675ebf4 | null | munish0838/mjlab-sprint2m-microduck-A2-r1 | [
{
"checkpoint": 250,
"sweep": [
{
"command_vx": 0.2,
"status": "fall",
"time_s": null,
"progress_m": 0.71
},
{
"command_vx": 0.25,
"status": "fall",
"time_s": null,
"progress_m": 0.66
},
{
"command_vx": 0.3... | true | 999 | 5.219 | 5 | 0.55 | ROUND 1 — run A2-r1, commit 4b675ebf4
hypothesis: —
change: no change (pristine recipe)
ARENA (local, Arena rules, platform: arm64 Darwin, arena 0.6.0)
checkpoint score finished best vx
250 50.000 0/5 0.50
500 7.739 5/5 0.50
750 5.399 5/5 0.55
999 5.21... | [] | null | null | null | null | null | null | null | null | null | 33 | arm64 Darwin | 0.6.0 |
C | C | 1 | null | null | null | null | null | null | null | null | null | null | null | null | null | null | f8a1f1cc5 | null | munish0838/mjlab-sprint2m-microduck-C-r1 | [
{
"checkpoint": 250,
"sweep": [
{
"command_vx": 0.2,
"status": "finish",
"time_s": 8.45,
"progress_m": null
},
{
"command_vx": 0.25,
"status": "fall",
"time_s": null,
"progress_m": 1.9
},
{
"command_vx": 0.... | true | 750 | 3.271 | 5 | 0.75 | ROUND 1 — run C-r1, commit f8a1f1cc5
hypothesis: —
change: forward_reward 2.0 -> 3.0; rerun of f8a1f1cc5
ARENA (local, Arena rules, platform: arm64 Darwin, arena 0.6.0)
checkpoint score finished best vx
250 6.584 5/5 0.50
500 3.582 5/5 0.70
750 3.271 5/5 0.75... | [] | null | null | null | null | null | null | null | null | null | 30 | arm64 Darwin | 0.6.0 |
C2 | C2 | 1 | null | null | null | null | null | null | null | null | null | null | null | null | null | null | ef4b93eb4 | null | munish0838/mjlab-sprint2m-microduck-C2-r1 | [
{
"checkpoint": 250,
"sweep": [
{
"command_vx": 0.2,
"status": "timeout",
"time_s": null,
"progress_m": 0.12
},
{
"command_vx": 0.25,
"status": "timeout",
"time_s": null,
"progress_m": 0.68
},
{
"command_vx... | true | 750 | 4.469 | 5 | 0.45 | ROUND 1 — run C2-r1, commit ef4b93eb4
hypothesis: —
change: forward_reward 2.0 -> 4.0; rerun of ef4b93eb4
ARENA (local, Arena rules, platform: arm64 Darwin, arena 0.6.0)
checkpoint score finished best vx
250 17.079 4/5 0.95
500 4.940 5/5 0.40
750 4.469 5/5 0.4... | [
"checkpoint 999 scored 21.932 s (3/5) after an earlier checkpoint had already reached 4.469 s — the run regressed, so its best may not reproduce on another seed"
] | null | null | null | null | null | null | null | null | null | 30 | arm64 Darwin | 0.6.0 |
C3 | C3 | 1 | null | null | null | null | null | null | null | null | null | null | null | null | null | null | aca704be2 | null | munish0838/mjlab-sprint2m-microduck-C3-r1 | [
{
"checkpoint": 250,
"sweep": [
{
"command_vx": 0.2,
"status": "finish",
"time_s": 7.9,
"progress_m": null
},
{
"command_vx": 0.25,
"status": "finish",
"time_s": 6.85,
"progress_m": null
},
{
"command_vx": ... | true | 999 | 3.159 | 5 | 0.7 | ROUND 1 — run C3-r1, commit aca704be2
hypothesis: —
change: forward_reward 2.0 -> 4.0; step_reward 3.0 -> 5.0; rerun of aca704be2
ARENA (local, Arena rules, platform: arm64 Darwin, arena 0.6.0)
checkpoint score finished best vx
250 5.958 5/5 0.90
500 3.875 5/5 0.85
750 ... | [
"checkpoint 750 scored 21.929 s (3/5) after an earlier checkpoint had already reached 3.875 s — the run regressed, so its best may not reproduce on another seed"
] | null | null | null | null | null | null | null | null | null | 48 | arm64 Darwin | 0.6.0 |
C4 | C4 | 1 | null | null | null | null | null | null | null | null | null | null | null | null | null | null | 18812a746 | null | munish0838/mjlab-sprint2m-microduck-C4-r1 | [
{
"checkpoint": 250,
"sweep": [
{
"command_vx": 0.2,
"status": "fall",
"time_s": null,
"progress_m": 1.01
},
{
"command_vx": 0.25,
"status": "fall",
"time_s": null,
"progress_m": 1.22
},
{
"command_vx": 0.3... | true | 999 | 3.085 | 5 | 0.8 | ROUND 1 — run C4-r1, commit 18812a746
hypothesis: —
change: forward_reward 2.0 -> 4.0; upright_reward 2.0 -> 3.0; step_reward 3.0 -> 5.0; rerun of 18812a746
ARENA (local, Arena rules, platform: arm64 Darwin, arena 0.6.0)
checkpoint score finished best vx
250 5.220 5/5 0.60
500 12.872 ... | [
"checkpoint 500 scored 12.872 s (4/5) after an earlier checkpoint had already reached 5.220 s — the run regressed, so its best may not reproduce on another seed"
] | null | null | null | null | null | null | null | null | null | 46 | arm64 Darwin | 0.6.0 |
C5 | C5 | 1 | null | null | null | null | null | null | null | null | null | null | null | null | null | null | 17df155b2 | null | munish0838/mjlab-sprint2m-microduck-C5-r1 | [
{
"checkpoint": 250,
"sweep": [
{
"command_vx": 0.2,
"status": "timeout",
"time_s": null,
"progress_m": 1.39
},
{
"command_vx": 0.25,
"status": "timeout",
"time_s": null,
"progress_m": 1.47
},
{
"command_vx... | true | 999 | 2.797 | 5 | 0.9 | ROUND 1 — run C5-r1, commit 17df155b2
hypothesis: —
change: forward_reward 2.0 -> 4.0; upright_reward 2.0 -> 3.0; step_reward 3.0 -> 5.0; rerun of 17df155b2
ARENA (local, Arena rules, platform: arm64 Darwin, arena 0.6.0)
checkpoint score finished best vx
250 14.408 4/5 0.50
500 6.306 ... | [] | null | null | null | null | null | null | null | null | null | 34 | arm64 Darwin | 0.6.0 |
C6 | C6 | 1 | null | null | null | null | null | null | null | null | null | null | null | null | null | null | d2975dcbf | null | munish0838/mjlab-sprint2m-microduck-C6-r1 | [
{
"checkpoint": 250,
"sweep": [
{
"command_vx": 0.2,
"status": "timeout",
"time_s": null,
"progress_m": 1.97
},
{
"command_vx": 0.25,
"status": "finish",
"time_s": 7.82,
"progress_m": null
},
{
"command_vx"... | true | 500 | 3.252 | 5 | 0.75 | ROUND 1 — run C6-r1, commit d2975dcbf
hypothesis: —
change: forward_reward 2.0 -> 4.0; upright_reward 2.0 -> 3.0; step_reward 3.0 -> 5.0; rerun of d2975dcbf
ARENA (local, Arena rules, platform: arm64 Darwin, arena 0.6.0)
checkpoint score finished best vx
250 24.456 3/5 0.25
500 3.252 ... | [
"checkpoint 750 scored 21.851 s (3/5) after an earlier checkpoint had already reached 3.252 s — the run regressed, so its best may not reproduce on another seed",
"checkpoint 999 scored 12.309 s (4/5) after an earlier checkpoint had already reached 3.252 s — the run regressed, so its best may not reproduce on ano... | null | null | null | null | null | null | null | null | null | 36 | arm64 Darwin | 0.6.0 |
C7 | C7 | 1 | null | null | null | null | null | null | null | null | null | null | null | null | null | null | 9dea21c90 | null | munish0838/mjlab-sprint2m-microduck-C7-r1 | [
{
"checkpoint": 250,
"sweep": [
{
"command_vx": 0.2,
"status": "finish",
"time_s": 7.41,
"progress_m": null
},
{
"command_vx": 0.25,
"status": "finish",
"time_s": 7.17,
"progress_m": null
},
{
"command_vx":... | true | 999 | 2.979 | 5 | 0.8 | ROUND 1 — run C7-r1, commit 9dea21c90
hypothesis: —
change: forward_reward 2.0 -> 4.0; upright_reward 2.0 -> 3.0; step_reward 3.0 -> 5.0; rerun of 9dea21c90
ARENA (local, Arena rules, platform: arm64 Darwin, arena 0.6.0)
checkpoint score finished best vx
250 16.159 4/5 0.70
500 22.279 ... | [] | null | null | null | null | null | null | null | null | null | 34 | arm64 Darwin | 0.6.0 |
S1 | S1 | 1 | You are an autonomous robot-learning engineer. Your job: make Pollen Robotics'
Microduck finish the Arena's 2 m Sprint faster, by changing its training recipe.
How this works:
- You never run commands. A human operator runs a fixed pipeline for you: it
applies your edits exactly as written, runs the repo's tests and... | # A2. Pollen's agent guide for this challenge (verbatim)
This is `src/microduck_challenges/sprint_2m/AGENTS.md`, unedited. It is the
guide Pollen wrote for coding agents on this challenge: follow its rules, its
do-not-touch list and its discipline.
# AGENTS.md: 2 m Sprint
You are helping a user train a Microduck pol... | 34,126 | No prior run exists in this session. The only measured bar is Pollen's reference at 4.883 s, 5/5 finishes, 0.55 m/s. The recipe already commands the full Arena sweep, 0.20–1.20 m/s straight ahead, with heading_hold and upright rewards in place. The first change should follow the knob-first discipline and increase the d... | Raising forward_reward from 2.0 to 3.0 makes track_linear_velocity more important relative to the gait and posture terms, so the policy should push farther up the speed sweep and finish faster while heading_hold keeps it straight. | knob | null | {
"forward_reward": 3,
"upright_reward": 2,
"step_reward": 3
} | [
{
"file": "sprint_2m/challenge.toml",
"old": "[params.forward_reward]\nlabel = \"Follow the speed order\"\nvalue = 2.0\nmin = 0.5\nmax = 5.0",
"new": "[params.forward_reward]\nlabel = \"Follow the speed order\"\nvalue = 3.0\nmin = 0.5\nmax = 5.0"
}
] | [] | {
"best_score_s": 4.4,
"finished": 5,
"best_vx": 0.6,
"what_to_watch": "fell_over should stay near 0; track_linear_velocity should rise and the best checkpoint should use a higher command speed than 0.55 m/s without increased wall contact or falls."
} | We need respond JSON only. Need decide first run change. Need reason carefully. We have no previous report. Budget 1000 iterations. Need beat reference 4.883 at 0.55. Need propose one change. Must follow discipline knob first. We can adjust knobs. Need choose likely improvement. Need inspect known: reference trained up... | diff --git a/src/microduck_challenges/sprint_2m/challenge.toml b/src/microduck_challenges/sprint_2m/challenge.toml
index 112fba7..b175d55 100644
--- a/src/microduck_challenges/sprint_2m/challenge.toml
+++ b/src/microduck_challenges/sprint_2m/challenge.toml
@@ -11,7 +11,7 @@ seed = 1
[params.forward_reward]... | null | f8a1f1cc5 | beam/S1/r1 | munish0838/mjlab-sprint2m-microduck-S1-r1 | [
{
"checkpoint": 250,
"sweep": [
{
"command_vx": 0.2,
"status": "fall",
"time_s": null,
"progress_m": 1.09
},
{
"command_vx": 0.25,
"status": "fall",
"time_s": null,
"progress_m": 1.15
},
{
"command_vx": 0.3... | true | 999 | 3.403 | 5 | 0.75 | ROUND 1 — run S1-r1, commit f8a1f1cc5
hypothesis: Raising forward_reward from 2.0 to 3.0 makes track_linear_velocity more important relative to the gait and posture terms, so the policy should push farther up the speed sweep and finish faster while heading_hold keeps it straight.
change: forward_reward 2.0 -> 3.0
AREN... | [] | 4.4 | -0.997 | 37,242 | 4,430 | 4,021 | 1 | Beam-1.0-501B-A23B | 33.2 | stop | 31 | arm64 Darwin | 0.6.0 |
S1 | S1 | 2 | You are an autonomous robot-learning engineer. Your job: make Pollen Robotics'
Microduck finish the Arena's 2 m Sprint faster, by changing its training recipe.
How this works:
- You never run commands. A human operator runs a fixed pipeline for you: it
applies your edits exactly as written, runs the repo's tests and... | # A2. Pollen's agent guide for this challenge (verbatim)
This is `src/microduck_challenges/sprint_2m/AGENTS.md`, unedited. It is the
guide Pollen wrote for coding agents on this challenge: follow its rules, its
do-not-touch list and its discipline.
# AGENTS.md: 2 m Sprint
You are helping a user train a Microduck pol... | 35,444 | Round 1 raised forward_reward 2.0→3.0 and produced a large, healthy gain: 3.403 s, 5/5, 0.75 m/s (vs control 5.438 s), with fell_over 0.167 and episode length 983. The sweep shows actual speed plateaus near 0.59–0.60 m/s above command 0.75 (0.80→3.52 s, 0.85→6.25 s), so the remaining bottleneck is the policy's willingn... | Raising forward_reward from 3.0 to 4.0 makes speed tracking dominate the gait/posture stack even more, pushing the speed plateau beyond ~0.60 m/s actual while heading_hold keeps drift bounded. | knob | null | {
"forward_reward": 4,
"upright_reward": 2,
"step_reward": 3
} | [] | [] | {
"best_score_s": 3.1,
"finished": 5,
"best_vx": 0.8,
"what_to_watch": "fell_over stays near 0; track_linear_velocity rises above 1.275; best checkpoint uses a higher command speed than 0.75 with average speed above ~0.60 m/s; no nan_state."
} | Let me analyze the situation carefully.
Round 1 result: raising forward_reward from 2.0 to 3.0 gave a big improvement: 3.403s, 5/5, 0.75 m/s. That's much faster than the reference (4.883s) and the control (5.438s). The sweep shows the policy walks well up to ~0.80-0.85 m/s, then degrades. Best vx 0.75.
Now I have 2 m... | diff --git a/src/microduck_challenges/sprint_2m/challenge.toml b/src/microduck_challenges/sprint_2m/challenge.toml
index 112fba7..0ccc076 100644
--- a/src/microduck_challenges/sprint_2m/challenge.toml
+++ b/src/microduck_challenges/sprint_2m/challenge.toml
@@ -11,7 +11,7 @@ seed = 1
[params.forward_reward]... | null | ef4b93eb4 | beam/S1/r2 | munish0838/mjlab-sprint2m-microduck-S1-r2 | [
{
"checkpoint": 250,
"sweep": [
{
"command_vx": 0.2,
"status": "timeout",
"time_s": null,
"progress_m": 1.28
},
{
"command_vx": 0.25,
"status": "finish",
"time_s": 6.89,
"progress_m": null
},
{
"command_vx"... | true | 750 | 2.795 | 5 | 0.9 | ROUND 2 — run S1-r2, commit ef4b93eb4
hypothesis: Raising forward_reward from 3.0 to 4.0 makes speed tracking dominate the gait/posture stack even more, pushing the speed plateau beyond ~0.60 m/s actual while heading_hold keeps drift bounded.
change: forward_reward 2.0 -> 4.0
ARENA (local, Arena rules, platform: arm64... | [
"checkpoint 999 scored 6.114 s (5/5) after an earlier checkpoint had already reached 2.795 s — the run regressed, so its best may not reproduce on another seed"
] | 3.1 | -0.305 | 39,435 | 4,458 | 4,090 | 1 | Beam-1.0-501B-A23B | 26.9 | stop | 31 | arm64 Darwin | 0.6.0 |
S1 | S1 | 3 | You are an autonomous robot-learning engineer. Your job: make Pollen Robotics'
Microduck finish the Arena's 2 m Sprint faster, by changing its training recipe.
How this works:
- You never run commands. A human operator runs a fixed pipeline for you: it
applies your edits exactly as written, runs the repo's tests and... | # A2. Pollen's agent guide for this challenge (verbatim)
This is `src/microduck_challenges/sprint_2m/AGENTS.md`, unedited. It is the
guide Pollen wrote for coding agents on this challenge: follow its rules, its
do-not-touch list and its discipline.
# AGENTS.md: 2 m Sprint
You are helping a user train a Microduck pol... | 35,507 | Round 2 improved the best score to 2.795 s, 5/5, at 0.90 m/s, but its sweep shows the policy still falls at 0.95–1.00 m/s and stalls at higher commands; the known failure mode is crab-walking/yaw drift at speed. Since every run resets to the pristine recipe, the forward_reward gain from round 2 must be carried forward ... | Keeping forward_reward at 4.0 and doubling heading_hold's weight from 1.0 to 2.0 makes sustained yaw drift at speed cost enough to discourage crab-walking while preserving the speed-tracking gain, so the policy should finish reliably at 0.95 m/s or above. | combined | The forward_reward knob from round 2 is not present in the pristine recipe, so it is re-applied together with the single new heading_hold weight change. | {
"forward_reward": 4,
"upright_reward": 2,
"step_reward": 3
} | [
{
"file": "sprint_2m/challenge.toml",
"old": "[params.forward_reward]\nlabel = \"Follow the speed order\"\nvalue = 2.0\nmin = 0.5\nmax = 5.0",
"new": "[params.forward_reward]\nlabel = \"Follow the speed order\"\nvalue = 4.0\nmin = 0.5\nmax = 5.0"
},
{
"file": "sprint_2m/env.py",
"old... | [] | {
"best_score_s": 2.5,
"finished": 5,
"best_vx": 0.95,
"what_to_watch": "heading_hold reward stays higher than round 2, fell_over stays near zero, the best checkpoint is at 750 or later, and 0.95 m/s finishes instead of falling."
} | We need respond JSON only. Need determine intended change. We need inspect issue: previous reply likely changed forward_reward 2.0 -> 4.0? But retry says edits old found 2 times because `value = 2.0\nmin = 0.5\nmax=5.0` appears for both forward_reward and upright_reward? Need unique quote. We need now propose exactly o... | diff --git a/src/microduck_challenges/sprint_2m/challenge.toml b/src/microduck_challenges/sprint_2m/challenge.toml
index 112fba7..0ccc076 100644
--- a/src/microduck_challenges/sprint_2m/challenge.toml
+++ b/src/microduck_challenges/sprint_2m/challenge.toml
@@ -11,7 +11,7 @@ seed = 1
[params.forward_reward]... | null | c8caebb8b | beam/S1/r3 | munish0838/mjlab-sprint2m-microduck-S1-r3 | [
{
"checkpoint": 250,
"sweep": [
{
"command_vx": 0.2,
"status": "fall",
"time_s": null,
"progress_m": 1.08
},
{
"command_vx": 0.25,
"status": "fall",
"time_s": null,
"progress_m": 1.11
},
{
"command_vx": 0.3... | true | 750 | 2.816 | 5 | 0.8 | ROUND 3 — run S1-r3, commit c8caebb8b
hypothesis: Keeping forward_reward at 4.0 and doubling heading_hold's weight from 1.0 to 2.0 makes sustained yaw drift at speed cost enough to discourage crab-walking while preserving the speed-tracking gain, so the policy should finish reliably at 0.95 m/s or above.
change: forwar... | [] | 2.5 | 0.316 | 39,760 | 7,150 | 6,530 | 2 | Beam-1.0-501B-A23B | 48.8 | stop | 31 | arm64 Darwin | 0.6.0 |
S1 | S1 | 4 | You are an autonomous robot-learning engineer. Your job: make Pollen Robotics'
Microduck finish the Arena's 2 m Sprint faster, by changing its training recipe.
How this works:
- You never run commands. A human operator runs a fixed pipeline for you: it
applies your edits exactly as written, runs the repo's tests and... | # A2. Pollen's agent guide for this challenge (verbatim)
This is `src/microduck_challenges/sprint_2m/AGENTS.md`, unedited. It is the
guide Pollen wrote for coding agents on this challenge: follow its rules, its
do-not-touch list and its discipline.
# AGENTS.md: 2 m Sprint
You are helping a user train a Microduck pol... | 35,764 | Round 3 (forward_reward 4.0 + heading_hold weight 2.0) scored 2.816 s, 5/5 at 0.80 m/s, slightly worse than round 2's 2.795 s at 0.90 m/s with heading_hold weight 1.0 — so extra heading weight hurt top speed rather than helping. The sweep shows a stability cliff between 0.90 (2.93 s) and 0.95 (4.24 s), and the diagnose... | Tightening heading_hold std from 0.4 to 0.2 at unchanged weight 1.0 makes sustained yaw drift at speed cost enough gradient to discourage crab-walking without over-constraining the gait, pushing the failure cliff above 0.90 m/s while keeping 5/5 finishes. | env | null | {
"forward_reward": 4,
"upright_reward": 2,
"step_reward": 3
} | [
{
"file": "sprint_2m/env.py",
"old": " cfg.rewards[\"heading_hold\"] = RewardTermCfg(\n func=microduck_mdp.heading_hold_reward,\n weight=1.0,\n params={\"std\": 0.4, \"asset_cfg\": SceneEntityCfg(\"robot\")},\n )",
"new": " cfg.rewards[\"heading_hold\"] = RewardTermCfg(\n ... | [] | {
"best_score_s": 2.6,
"finished": 5,
"best_vx": 0.95,
"what_to_watch": "heading_hold reward should stay high with low variance, fell_over near zero, best checkpoint at 750 or later, and the sweep cliff should move from 0.90 to 0.95+ m/s; if std is too tight and air_time collapses or entropy drops, the gait ove... | Let me analyze the situation carefully.
This is round 4 of 4 in the session. The budget is 1000 iterations, 4096 envs, seed 1. Let me review the history:
Round 1: forward_reward 2.0→3.0, best ckpt 999, 3.403s, 5/5, 0.75 m/s
Round 2: forward_reward 2.0→4.0, best ckpt 750, 2.795s, 5/5, 0.90 m/s
Round 3: forward_reward ... | diff --git a/src/microduck_challenges/sprint_2m/challenge.toml b/src/microduck_challenges/sprint_2m/challenge.toml
index 112fba7..0ccc076 100644
--- a/src/microduck_challenges/sprint_2m/challenge.toml
+++ b/src/microduck_challenges/sprint_2m/challenge.toml
@@ -11,7 +11,7 @@ seed = 1
[params.forward_reward]... | null | f0c809396 | beam/S1/r4 | munish0838/mjlab-sprint2m-microduck-S1-r4 | [
{
"checkpoint": 250,
"sweep": [
{
"command_vx": 0.2,
"status": "fall",
"time_s": null,
"progress_m": 0.79
},
{
"command_vx": 0.25,
"status": "fall",
"time_s": null,
"progress_m": 1.24
},
{
"command_vx": 0.3... | true | 999 | 3.367 | 5 | 0.9 | ROUND 4 — run S1-r4, commit f0c809396
hypothesis: Tightening heading_hold std from 0.4 to 0.2 at unchanged weight 1.0 makes sustained yaw drift at speed cost enough gradient to discourage crab-walking without over-constraining the gait, pushing the failure cliff above 0.90 m/s while keeping 5/5 finishes.
change: forwar... | [
"checkpoint 750 scored 22.298 s (3/5) after an earlier checkpoint had already reached 3.470 s — the run regressed, so its best may not reproduce on another seed"
] | 2.6 | 0.767 | 39,908 | 5,546 | 4,902 | 1 | Beam-1.0-501B-A23B | 31.6 | stop | 31 | arm64 Darwin | 0.6.0 |
S1 | S1 | 5 | You are an autonomous robot-learning engineer. Your job: make Pollen Robotics'
Microduck finish the Arena's 2 m Sprint faster, by changing its training recipe.
How this works:
- You never run commands. A human operator runs a fixed pipeline for you: it
applies your edits exactly as written, runs the repo's tests and... | # A2. Pollen's agent guide for this challenge (verbatim)
This is `src/microduck_challenges/sprint_2m/AGENTS.md`, unedited. It is the
guide Pollen wrote for coding agents on this challenge: follow its rules, its
do-not-touch list and its discipline.
# AGENTS.md: 2 m Sprint
You are helping a user train a Microduck pol... | 35,950 | Round 4 (forward_reward 4.0 + heading_hold std 0.2) scored 3.367 s, 5/5 at 0.90 m/s — worse than round 2 (2.795 s, 5/5 at 0.90 m/s) which used default heading_hold std 0.4. Tightening the heading_hold std over-constrained the gait and hurt stability, as the sweep showed scattered falls across many speeds. The best conf... | Raising step_reward from 3.0 to 5.0 increases air_time reward weight, encouraging more dynamic and stable stepping at high commanded speeds while forward_reward 4.0 maintains speed tracking, pushing the failure cliff above 0.90 m/s with 5/5 finishes. | knob | null | {
"forward_reward": 4,
"upright_reward": 2,
"step_reward": 5
} | [] | [] | {
"best_score_s": 2.6,
"finished": 5,
"best_vx": 0.95,
"what_to_watch": "air_time reward should increase relative to round 2, fell_over should stay near zero, heading_hold should remain at default levels, and the sweep cliff should move from 0.90 to 0.95+ m/s; if step_reward is too high and the policy marches i... | Let me analyze the situation carefully.
We're on round 5 of 6. The goal is to make the Microduck finish the 2m sprint faster.
Current best results:
- Round 1: forward_reward 3.0 → 3.403s, 5/5, 0.75 m/s
- Round 2: forward_reward 4.0 → 2.795s, 5/5, 0.90 m/s
- Round 3: forward_reward 4.0 + heading_hold weight 2.0 → 2.81... | diff --git a/src/microduck_challenges/sprint_2m/challenge.toml b/src/microduck_challenges/sprint_2m/challenge.toml
index 112fba7..a47d8ae 100644
--- a/src/microduck_challenges/sprint_2m/challenge.toml
+++ b/src/microduck_challenges/sprint_2m/challenge.toml
@@ -11,7 +11,7 @@ seed = 1
[params.forward_reward]... | null | aca704be2 | beam/S1/r5 | munish0838/mjlab-sprint2m-microduck-S1-r5 | [
{
"checkpoint": 250,
"sweep": [
{
"command_vx": 0.2,
"status": "timeout",
"time_s": null,
"progress_m": 0.87
},
{
"command_vx": 0.25,
"status": "timeout",
"time_s": null,
"progress_m": 1.02
},
{
"command_vx... | true | 999 | 2.751 | 5 | 0.8 | ROUND 5 — run S1-r5, commit aca704be2
hypothesis: Raising step_reward from 3.0 to 5.0 increases air_time reward weight, encouraging more dynamic and stable stepping at high commanded speeds while forward_reward 4.0 maintains speed tracking, pushing the failure cliff above 0.90 m/s with 5/5 finishes.
change: forward_rew... | [] | 2.6 | 0.151 | 40,171 | 4,401 | 3,939 | 1 | Beam-1.0-501B-A23B | 22.8 | stop | 50 | arm64 Darwin | 0.6.0 |
S1 | S1 | 6 | You are an autonomous robot-learning engineer. Your job: make Pollen Robotics'
Microduck finish the Arena's 2 m Sprint faster, by changing its training recipe.
How this works:
- You never run commands. A human operator runs a fixed pipeline for you: it
applies your edits exactly as written, runs the repo's tests and... | # A2. Pollen's agent guide for this challenge (verbatim)
This is `src/microduck_challenges/sprint_2m/AGENTS.md`, unedited. It is the
guide Pollen wrote for coding agents on this challenge: follow its rules, its
do-not-touch list and its discipline.
# AGENTS.md: 2 m Sprint
You are helping a user train a Microduck pol... | 35,877 | Round 5 scored 2.751 s, 5/5 at 0.80 m/s (best so far), with forward_reward=4.0 and step_reward=5.0. Its sweep shows a fall band at 0.95-1.10 m/s (falls at 0.95, 1.00, 1.05, 1.10) but success at 0.85 (2.71 s) and 1.15-1.20, indicating trunk/yaw instability in the mid-speed range. upright_reward has never been changed fr... | Raising upright_reward from 2.0 to 3.0 strengthens trunk-verticality pressure at speed, reducing forward-lean falls in the 0.95-1.10 m/s band while preserving the dynamic gait from step_reward=5.0, pushing 5/5 robustness to 0.90+ m/s. | knob | null | {
"forward_reward": 4,
"upright_reward": 3,
"step_reward": 5
} | [] | [] | {
"best_score_s": 2.6,
"finished": 5,
"best_vx": 0.85,
"what_to_watch": "fell_over should drop below 0.25, heading_hold should stay near 0.25, and the 0.95-1.10 fall band should shrink; if upright is too strong, air_time or track_linear_velocity may drop."
} | Let me analyze the situation carefully.
We're in round 6 of 6. The last round (round 5) achieved 2.751 s, 5/5 at 0.80 m/s with knobs forward_reward=4.0, step_reward=5.0, upright_reward=2.0 (default).
The best so far is round 2: 2.795 s at 0.90 m/s... wait, round 5 scored 2.751 s at 0.80 m/s. Lower time is better. So ... | diff --git a/src/microduck_challenges/sprint_2m/challenge.toml b/src/microduck_challenges/sprint_2m/challenge.toml
index 112fba7..1c46722 100644
--- a/src/microduck_challenges/sprint_2m/challenge.toml
+++ b/src/microduck_challenges/sprint_2m/challenge.toml
@@ -11,18 +11,18 @@ seed = 1
[params.forward_rewar... | null | 18812a746 | beam/S1/r6 | munish0838/mjlab-sprint2m-microduck-S1-r6 | [
{
"checkpoint": 250,
"sweep": [
{
"command_vx": 0.2,
"status": "timeout",
"time_s": null,
"progress_m": -0.27
},
{
"command_vx": 0.25,
"status": "fall",
"time_s": null,
"progress_m": 0.39
},
{
"command_vx":... | true | 750 | 2.945 | 5 | 0.9 | ROUND 6 — run S1-r6, commit 18812a746
hypothesis: Raising upright_reward from 2.0 to 3.0 strengthens trunk-verticality pressure at speed, reducing forward-lean falls in the 0.95-1.10 m/s band while preserving the dynamic gait from step_reward=5.0, pushing 5/5 robustness to 0.90+ m/s.
change: forward_reward 2.0 -> 4.0; ... | [
"checkpoint 500 scored 21.963 s (3/5) after an earlier checkpoint had already reached 5.264 s — the run regressed, so its best may not reproduce on another seed"
] | 2.6 | 0.345 | 40,121 | 6,435 | 5,980 | 1 | Beam-1.0-501B-A23B | 26.3 | stop | 48 | arm64 Darwin | 0.6.0 |
S1 | S1 | 7 | You are an autonomous robot-learning engineer. Your job: make Pollen Robotics'
Microduck finish the Arena's 2 m Sprint faster, by changing its training recipe.
How this works:
- You never run commands. A human operator runs a fixed pipeline for you: it
applies your edits exactly as written, runs the repo's tests and... | # A2. Pollen's agent guide for this challenge (verbatim)
This is `src/microduck_challenges/sprint_2m/AGENTS.md`, unedited. It is the
guide Pollen wrote for coding agents on this challenge: follow its rules, its
do-not-touch list and its discipline.
# AGENTS.md: 2 m Sprint
You are helping a user train a Microduck pol... | 36,306 | Round 6 (forward 4.0, upright 3.0, step 5.0) scored 2.945 s, 5/5 at 0.90 m/s (seed 1) and 3.085 s (seed 2), spread 0.140 s — stable but not faster than round 5's 2.751 s. The sweep shows a failure cliff at 1.00–1.05 m/s (falls at 1.76 m and 1.57 m), while 0.80–0.90 m/s is the reliable sweet spot. Critically, checkpoint... | Halving the action_rate curriculum's final weights (−0.1→−0.4 instead of −0.1→−0.6 by iter 1000) reduces smoothing pressure on the dynamic gait, allowing the policy to sustain speeds beyond 0.90 m/s and pushing the failure cliff above 1.00 m/s while keeping 5/5 finishes. | env | null | {
"forward_reward": 4,
"upright_reward": 3,
"step_reward": 5
} | [
{
"file": "sprint_2m/env.py",
"old": " \"weight_stages\": [\n {\"step\": 0, \"weight\": -0.1},\n {\"step\": 500 * NUM_STEPS_PER_ENV, \"weight\": -0.2},\n {\"step\": 750 * NUM_STEPS_PER_ENV, \"weight\": -0.4},\n {\"step\": 1000 * NUM_STEP... | [] | {
"best_score_s": 2.8,
"finished": 5,
"best_vx": 0.9,
"what_to_watch": "Whether ckpt 999 now beats ckpt 750 (if action_rate was the bottleneck, the best checkpoint should shift to 999); whether the 1.00–1.05 m/s fall band shrinks or disappears; whether fell_over stays below 0.15 and heading_hold stays above 0.2... | Let me analyze the situation carefully.
**Current state (Round 6 results):**
- Best recipe so far: forward_reward=4.0, upright_reward=3.0, step_reward=5.0
- Best checkpoint: 750, 2.945 s, 5/5 at 0.90 m/s (seed 1), 3.085 s seed 2, spread 0.140 s
- Reference: 4.883 s at 0.55 m/s; control: 5.438 s at 0.50 m/s
**Key obse... | diff --git a/src/microduck_challenges/sprint_2m/challenge.toml b/src/microduck_challenges/sprint_2m/challenge.toml
index 112fba7..1c46722 100644
--- a/src/microduck_challenges/sprint_2m/challenge.toml
+++ b/src/microduck_challenges/sprint_2m/challenge.toml
@@ -11,18 +11,18 @@ seed = 1
[params.forward_rewar... | null | 17df155b2 | beam/S1/r7 | munish0838/mjlab-sprint2m-microduck-S1-r7 | [
{
"checkpoint": 250,
"sweep": [
{
"command_vx": 0.2,
"status": "finish",
"time_s": 25.21,
"progress_m": null
},
{
"command_vx": 0.25,
"status": "timeout",
"time_s": null,
"progress_m": 1.39
},
{
"command_vx... | true | 750 | 2.917 | 5 | 0.85 | ROUND 7 — run S1-r7, commit 17df155b2
hypothesis: Halving the action_rate curriculum's final weights (−0.1→−0.4 instead of −0.1→−0.6 by iter 1000) reduces smoothing pressure on the dynamic gait, allowing the policy to sustain speeds beyond 0.90 m/s and pushing the failure cliff above 1.00 m/s while keeping 5/5 finishes... | [] | 2.8 | 0.117 | 40,788 | 10,221 | 9,256 | 1 | Beam-1.0-501B-A23B | 48.2 | stop | 36 | arm64 Darwin | 0.6.0 |
S1 | S1 | 8 | You are an autonomous robot-learning engineer. Your job: make Pollen Robotics'
Microduck finish the Arena's 2 m Sprint faster, by changing its training recipe.
How this works:
- You never run commands. A human operator runs a fixed pipeline for you: it
applies your edits exactly as written, runs the repo's tests and... | # A2. Pollen's agent guide for this challenge (verbatim)
This is `src/microduck_challenges/sprint_2m/AGENTS.md`, unedited. It is the
guide Pollen wrote for coding agents on this challenge: follow its rules, its
do-not-touch list and its discipline.
# AGENTS.md: 2 m Sprint
You are helping a user train a Microduck pol... | 36,657 | Round 7 reproduced well across seeds (2.917 s and 2.797 s, spread 0.120 s) and improved robustness, but the sweep still falls at 0.95–1.00 m/s, likely from sustained yaw drift/crab-walking at speed. Heading hold currently charges only a tiny penalty at a 3° offset because std=0.4, so tightening it to 0.2 should discour... | Keeping the round-7 recipe and tightening heading_hold std from 0.4 to 0.2 makes sustained yaw drift at speed costly enough to reduce crab-walking, pushing the 0.95–1.00 m/s failure band above 1.00 m/s while keeping 5/5 finishes. | combined | The winning knobs and action-rate curriculum from round 7 must be carried into the pristine recipe; the single new change is tightening heading_hold std from 0.4 to 0.2. | {
"forward_reward": 4,
"upright_reward": 3,
"step_reward": 5
} | [
{
"file": "sprint_2m/challenge.toml",
"old": "[params.forward_reward]\nlabel = \"Follow the speed order\"\nvalue = 2.0",
"new": "[params.forward_reward]\nlabel = \"Follow the speed order\"\nvalue = 4.0"
},
{
"file": "sprint_2m/challenge.toml",
"old": "[params.upright_reward]\nlabel = \"Stay ... | [] | {
"best_score_s": 2.7,
"finished": 5,
"best_vx": 0.9,
"what_to_watch": "Whether checkpoint 999 beats 750, whether the 0.95–1.00 m/s fall band shrinks, whether fell_over stays low and heading_hold stays above 0.2 at iter 999, and whether the two-seed spread stays below 0.25 s."
} | We need respond JSON one. Need decide next change. Need analyze deeply. We have 8 rounds. Current best recipe r7: knobs FR4, UR3, SR5, action_rate halved. Best 2.917s seed1 ckpt750, seed2 2.797 ckpt999 spread 0.12. Need beat. Need one change pristine. Need propose likely. Need consider known patterns: forward reward 4 ... | diff --git a/src/microduck_challenges/sprint_2m/challenge.toml b/src/microduck_challenges/sprint_2m/challenge.toml
index 112fba7..1c46722 100644
--- a/src/microduck_challenges/sprint_2m/challenge.toml
+++ b/src/microduck_challenges/sprint_2m/challenge.toml
@@ -11,18 +11,18 @@ seed = 1
[params.forward_rewar... | null | d2975dcbf | beam/S1/r8 | munish0838/mjlab-sprint2m-microduck-S1-r8 | [
{
"checkpoint": 250,
"sweep": [
{
"command_vx": 0.2,
"status": "fall",
"time_s": null,
"progress_m": 0.92
},
{
"command_vx": 0.25,
"status": "fall",
"time_s": null,
"progress_m": 0.92
},
{
"command_vx": 0.3... | true | 999 | 2.745 | 5 | 0.85 | ROUND 8 — run S1-r8, commit d2975dcbf
hypothesis: Keeping the round-7 recipe and tightening heading_hold std from 0.4 to 0.2 makes sustained yaw drift at speed costly enough to reduce crab-walking, pushing the 0.95–1.00 m/s failure band above 1.00 m/s while keeping 5/5 finishes.
change: forward_reward 2.0 -> 4.0; uprig... | [] | 2.7 | 0.045 | 41,320 | 6,833 | 5,891 | 1 | Beam-1.0-501B-A23B | 40.7 | stop | 34 | arm64 Darwin | 0.6.0 |
S1 | S1 | 9 | You are an autonomous robot-learning engineer. Your job: make Pollen Robotics'
Microduck finish the Arena's 2 m Sprint faster, by changing its training recipe.
How this works:
- You never run commands. A human operator runs a fixed pipeline for you: it
applies your edits exactly as written, runs the repo's tests and... | # A2. Pollen's agent guide for this challenge (verbatim)
This is `src/microduck_challenges/sprint_2m/AGENTS.md`, unedited. It is the
guide Pollen wrote for coding agents on this challenge: follow its rules, its
do-not-touch list and its discipline.
# AGENTS.md: 2 m Sprint
You are helping a user train a Microduck pol... | 36,912 | Round 8 improved seed-1 to 2.745 s but destroyed reproducibility (spread 0.507 s, seed 2 regressed to 3.252 s at ckpt 500 and collapsed at 750/999). The culprit was tightening heading_hold std from 0.4 to 0.2, which makes the reward landscape too sharp and seed-sensitive. Round 7 (std 0.4, halved action_rate) was the b... | Doubling heading_hold weight from 1.0 to 2.0 on top of the round-7 recipe increases yaw-drift gradient at speed to reduce crab-walking, while keeping std at 0.4 preserves the tight two-seed spread, pushing best finish above 0.85 m/s at 5/5. | env | null | {
"forward_reward": 4,
"upright_reward": 3,
"step_reward": 5
} | [
{
"file": "sprint_2m/env.py",
"old": " cfg.rewards[\"heading_hold\"] = RewardTermCfg(\n func=microduck_mdp.heading_hold_reward,\n weight=1.0,\n params={\"std\": 0.4, \"asset_cfg\": SceneEntityCfg(\"robot\")},\n )",
"new": " cfg.rewards[\"heading_hold\"] = RewardTermCfg(\n ... | [] | {
"best_score_s": 2.8,
"finished": 5,
"best_vx": 0.85,
"what_to_watch": "Whether heading_hold contribution doubles and the 0.90-1.00 m/s band finishes reliably, whether ckpt 999 beats 750, whether fell_over stays near 0.3 or lower, and whether the two-seed spread stays below 0.25 s (unlike round 8's 0.507 s)."
... | Let me analyze the situation carefully.
We're on round 9 of 10. The budget is 1000 iterations. The best so far has been round 8: 2.745 s, 5/5 at 0.85 m/s, but it did NOT reproduce on seed 2 (spread 0.507 s, seed 2 best was 3.252 s at checkpoint 500, and regressed badly at later checkpoints).
Key issues from the histo... | diff --git a/src/microduck_challenges/sprint_2m/challenge.toml b/src/microduck_challenges/sprint_2m/challenge.toml
index 112fba7..1c46722 100644
--- a/src/microduck_challenges/sprint_2m/challenge.toml
+++ b/src/microduck_challenges/sprint_2m/challenge.toml
@@ -11,18 +11,18 @@ seed = 1
[params.forward_rewar... | null | 9dea21c90 | beam/S1/r9 | munish0838/mjlab-sprint2m-microduck-S1-r9 | [
{
"checkpoint": 250,
"sweep": [
{
"command_vx": 0.2,
"status": "finish",
"time_s": 6.49,
"progress_m": null
},
{
"command_vx": 0.25,
"status": "fall",
"time_s": null,
"progress_m": 1.55
},
{
"command_vx": 0... | true | 750 | 3.394 | 5 | 0.75 | ROUND 9 — run S1-r9, commit 9dea21c90
hypothesis: Doubling heading_hold weight from 1.0 to 2.0 on top of the round-7 recipe increases yaw-drift gradient at speed to reduce crab-walking, while keeping std at 0.4 preserves the tight two-seed spread, pushing best finish above 0.85 m/s at 5/5.
change: forward_reward 2.0 ->... | [
"checkpoint 500 scored 12.904 s (4/5) after an earlier checkpoint had already reached 5.972 s — the run regressed, so its best may not reproduce on another seed"
] | 2.8 | 0.594 | 41,694 | 7,576 | 6,723 | 1 | Beam-1.0-501B-A23B | 34.9 | stop | 34 | arm64 Darwin | 0.6.0 |
S2 | S2 | 1 | You are an autonomous robot-learning engineer. Your job: make Pollen Robotics'
Microduck finish the Arena's 2 m Sprint faster, by changing its training recipe.
How this works:
- You never run commands. A human operator runs a fixed pipeline for you: it
applies your edits exactly as written, runs the repo's tests and... | # A2. Pollen's agent guide for this challenge (verbatim)
This is `src/microduck_challenges/sprint_2m/AGENTS.md`, unedited. It is the
guide Pollen wrote for coding agents on this challenge: follow its rules, its
do-not-touch list and its discipline.
# AGENTS.md: 2 m Sprint
You are helping a user train a Microduck pol... | 34,173 | No prior session runs exist. The unchanged 1000-iteration control scored 5.438 s with 5/5 finishes at 0.50 m/s (best checkpoint 750), versus Pollen's 4.883 s reference at 0.55 m/s. The speed-tracking term is currently only weighted 2.0, equal to or below stepping and upright rewards, so the first knob change should mak... | Raise forward_reward from 2.0 to 4.0 so tracking the commanded sprint speed carries more weight than stepping and upright regularizers, expecting a faster best checkpoint while keeping 5/5 finishes. | knob | null | {
"forward_reward": 4,
"upright_reward": 2,
"step_reward": 3
} | [
{
"file": "sprint_2m/challenge.toml",
"old": "[params.forward_reward]\nlabel = \"Follow the speed order\"\nvalue = 2.0\nmin = 0.5\nmax = 5.0",
"new": "[params.forward_reward]\nlabel = \"Follow the speed order\"\nvalue = 4.0\nmin = 0.5\nmax = 5.0"
}
] | [] | {
"best_score_s": 5,
"finished": 5,
"best_vx": 0.6,
"what_to_watch": "fell_over should stay near 0, episode length should remain high, track_linear_velocity should grow, and the best checkpoint should move faster than the 0.50 m/s control without losing finishes."
} | We need solve optimization challenge. Need produce one JSON proposal for run 1 of 3. Need carefully analyze task, budget 1000 iters. Need choose change. We have no prior run in session except control metrics provided: unchanged recipe at 1000 iters scored 5.438 s, 5/5 at 0.50 m/s at checkpoint 750. Need beat. Reference... | diff --git a/src/microduck_challenges/sprint_2m/challenge.toml b/src/microduck_challenges/sprint_2m/challenge.toml
index 112fba7..0ccc076 100644
--- a/src/microduck_challenges/sprint_2m/challenge.toml
+++ b/src/microduck_challenges/sprint_2m/challenge.toml
@@ -11,7 +11,7 @@ seed = 1
[params.forward_reward]... | null | null | null | null | null | false | null | null | null | null | null | null | null | null | null | null | 3,912 | 1 | Beam-1.0-501B-A23B | 31.3 | stop | null | null | null |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.