Kang, Park et al. Agile perceptive multi-skill locomotion for quadrupedal robots in the wild. Science Robotics 2026. No official code. Written from scratch.
APT-RL doesn't let the policy drive the joints directly. It optimises thousands of trot and bound cycles offline, compresses them into a small latent space, and makes the policy act in that space: pick a gait twice a second, steer the latent, add small corrections. I rebuilt the whole chain for a 14.5 kg A1, which is a much weaker robot than the paper's. In simulation the A1 now walks on the prior: switch the prior off and the walking stops. In October it got its eyes, a real trot and a library that reaches 3 m/s. The best teacher so far succeeds on 83.5% of the paper’s obstacle courses in simulation. But it picks trot almost every time, and it still won’t jump a gap.
Most locomotion policies output joint position targets and learn gaits from scratch, shaped by reward terms. APT-RL moves the prior into the action space instead. First, a trajectory optimiser generates a large set of physically consistent gait cycles. It uses a single-rigid-body model (SRBD): the robot is treated as one rigid body pushed by ground forces, in the sagittal plane only. The paper's set is 180,000 trajectories, 90,000 trot and 90,000 bound, from −2 to 7 m/s. That's about 15.5 hours of motion, generated in 8 minutes.
A transformer VAE is trained on that set. Its encoder maps a short window of robot state to a 16-dimensional latent. Separate decoders, one per gait, map the latent back to joint torques. After training, the encoder is thrown away and the decoders are frozen. The RL policy then outputs three things. A gait bit, chosen at 2 Hz (trot below 0.5, bound above). A 16-dim latent and 12 auxiliary joint offsets, both at 100 Hz. What reaches the motors is the decoder's torque plus a PD term around the default pose, shifted by the auxiliary action:
τ = Dec_gait(z) + kp·(q_default − q + a_scale·a_aux) − kd·q̇
The prior shapes which motions are cheap, not which are allowed; whatever the 2D library can't express, like hip abduction or a jump, comes from the auxiliary channel. The paper's key check zeroes the decoder torque in a trained policy: success drops from 94.6% to 2.9%. I call that the ff=0 test and ran it on every policy.
The paper also has a perception half: depth plus a 2D lidar distilled from privileged height maps into a student policy. I haven't got there. Everything on this page is about the action space.
There's no official code, so the library half is all mine: the SRBD generator, the A1 feasibility filter, the TVAE and the MuJoCo replay harness. Where the paper only said "following [59]", I filled in the details from the two MIT Cheetah 2 papers it defers to. The A1 constants (14.5 kg all-in, gains, joint order) came from measurements on my own robot in the parkour project.
The RL side went through three bases, and they overlap in time:
Then I stopped for seven weeks to work on DreamWaQ, HIMLoco, PIE and MOVE. I came back on Sep 29 with a second read of the paper, and both repos moved together from there: library fixes in apt-a1, RL in apt-rl.
The paper's optimiser applies no torque, friction or joint limits; it leaves feasibility to RL and a powerful robot. I kept its method and added a filter against the A1's measured limits. Getting the generator right took a day of silent bugs:
Jᵀf, but the actuator torque is −Jᵀf. The filter only checked |τ|, so it never noticed. Replaying the library in MuJoCo caught it: the body sank faster than free fall, then got thrown to 0.56 m and flipped.Replay also showed that at 1.75–2.05 m/s the stored torques alone hit the 33.5 N·m ceiling in 92% of runs, leaving nothing for feedback. So the filter now keeps 25% of the torque budget in reserve. The corrected library, lib_v2, holds 47,174 trajectories: 19,389 trot and 27,785 bound.
A latent-size sweep showed four dimensions are enough for this library: torque fit was R² 0.9943 at both 4 and 16. That fits a library set by about six numbers. I kept 16 to match the paper.
On Sep 29 an audit showed I'd trained it wrong. I had fitted encoder and decoders jointly on torque. The paper trains in two stages: first a VAE on state alone, then decoders on top of the frozen encoder. Refitting that way gave R² 0.9685, with 9 active latent dimensions spread like a unit Gaussian, and that mattered a lot for RL. Two later library fixes came from RL failures: a four-contact trot (library_v3, Sep 30) and a fast bound to 3 m/s (library_v4, Oct 1).
The first 36 hours were mostly about how many ways a training curve can lie.
In MuJoCo, at 78.6M steps, the library policy beat a no-prior policy on speed tracking by 35%. Then I measured where the torque came from: the auxiliary channel was supplying 94% of it. Against a fairer control, a frozen random decoder with the same authority, the library lost: 0.482 m/s mean error against 0.283. I retracted the claim the same morning.
On the parkour stack, four of five motion-library runs collapsed to zero reward for good. The stack clips each step's reward at zero, and the paper's gait-diversity term at −15 costs up to 0.075 per step against at most 0.042 of positive reward. Reward was exactly zero, whatever the action. At −1 it trained, but finished at terrain level 1.885 against 2.103 for plain position control, with the decoder down to 15% of the torque.
That was my reaction to the unclipped −15 run, right after I had presented its 44.5% gait mixing as evidence that the architecture worked. It was mixing gaits at 0.006 m/s. The gait bit can be flipped without moving. Next video: "it walked to the edge of the gap and just stood there." The foot-collision term read height in world coordinates, so a foot standing on a 20 cm step drew the largest penalty in the table.
Since the last config that walked, I had changed seven things. Six configurations in a row stood still at 0.004–0.056 m/s, and none of it could be attributed. I pinned the working config back and kept only the three real bug fixes. That walked at 0.537 m/s against a 0.55 command (98% tracking), versus 56% for the baseline at the same commands. The baseline still climbed about 9% more terrain.
I had published the overnight report with a clip captioned as walking. "it doesn't walk. it just face plants and stays there." It was the wrong checkpoint rendered under the wrong play task, and I hadn't watched it. Since then, no clip goes out unless I've read its frames.
A contact probe showed the walker matched no library gait: no leg pair's stance overlap above 0.47, an arrhythmic crawl. Gait diversity back at −15 woke the selector (p(bound) 0.568), at the cost of 80% tracking and a near-dragging rear leg.
The paper anneals its gait terms over 50,000 epochs; my runs were 1,500. A 15,000-iteration run kept collapsing and recovering. Then: "screw it, lets focus on the fresh new implementation... free both gpus."
apt-rl starts from Isaac Lab's velocity task, which walks an A1 in minutes, with the motion-library action and the paper's observation, reward and PPO tables swapped in. An equation audit from page renders found eight mismatches in my first port; one, an action penalty summed over all 29 dimensions instead of 12, broke training on its own.
The shakedown runs, SD1 onward, kept failing the same way. Under the paper's weights, living costs more than it earns, so the policy finds the cheapest way out. That was death (SD1: Isaac Lab multiplies rewards by dt, so the −1.5 death penalty actually delivered −0.015), a joint-limit exit, kneeling, or standing still.
Open loop, no policy, latents driven as sinusoids: 0.004 m/s and zero contact changes. What matters is decoder torque against PD stiffness, roughly 0.5–1 rad of deflection on HOUND and 0.05 rad on the A1. Dropping kp to 20 and tripling decoder gain walked 0.511 m/s open loop. Under PPO, the policy then steered the latent to where the decoder goes quiet, and froze it in time. A stock Isaac Lab A1 task reached 0.91 tracking by iteration 205. The same task with only my action space swapped in reached 0.14.
Random-phase sinusoids only rocked the body. Real latents from 24 library trot cycles stepped all four feet at trot cadence. I cloned the actor from those, then found PPO was paid to take the gait apart: here the gait is action motion, and an action-rate penalty the paper doesn't have charged for it. The fourth run, bankwarm4, moved forward at iteration 400.
It wasn't one. Three legs at duty 0.72–0.76 and one at 0.30. The latent was pinned at its clamp, the decoder was putting out 181 N·m per step (ten times library level), and it fell twice in 500 steps. I downgraded the claim to "locomotes" and wrote a walk bar into the diagnostics: all duties 0.45–0.75, zero falls, diagonal sync above lateral by 0.15, decoder torque under 36 N·m per step. No APT-RL checkpoint has passed it.
A 0.35 rad error in the default pose had been cancelling the prior's stance torque. Once I fixed it, I ran matched arms. Without the prior: tracking 0.267 at iteration 200, the best cold learner yet. With it: 0.005, falling constantly. I concluded the prior was a disturbance, not a scaffold, on this robot.
The same day, on the side branch, I put the library action and the pose fix onto the parkour scaffold with its own rewards and clip. The prior arm kept pace with a no-prior control (terrain 4.24 at iteration 951, against 4.40 at 901) and by 600 stepped with all duties in band. A later no-prior arm there passed the duty bar while skimming, with swing peaks of 2–33 mm, so I retracted it and added a 40 mm swing-height check. That branch became the rear-drag investigation on the Extreme Parkour teacher.
On Sep 29 I reread the paper and the Aug 9 runs. Moving the spawn pose to the library's crouch had kept the 0.42 m spawn height, so every "with prior" episode began with a 15 cm drop; its 35–67-step episodes were just the time to fall. It had also run only 200–260 iterations.
The bigger finding was a framework default. The paper trained on legged_gym, where only_positive_rewards clips the summed reward at zero and adds the termination penalty after the clip. Isaac Lab never clips. That difference accounts for the net-negative early life in August. So I turned the clip on, ramped penalties in over 1,000 iterations, and moved to a deploy-matched actuator (kp 37, kd 0.9, gains randomised over 34–40, 5–10 ms delay). The policy moved, but by slamming the latent between its limits, and stood once the smoothness penalty reached full weight. A low-pass filter on the latent paid that bill; the joint-acceleration penalty took over.
What finally worked was the paper-spec two-stage VAE plus a second audit's fixes. Those included the PPO regularisers, which had been running as an extra optimiser step outside the loss. I also defined gates, sized against the paper's own training curve:
| Gate | Pass criterion (sim) | Result |
|---|---|---|
| G1 no retreat | Tracking not falling after the penalty ramp completes | Pass, Sep 30 |
| G2 walker | Success ≥ 0.8, tracking ≥ 0.5, ff=0 tracking under a third | Pass, Sep 30 (bound only) |
| G3 gait use | Both forced gaits ≥ 0.7; p(bound) rises with speed | Fail, Oct 1 |
| G4 terrain teacher | Curriculum level ≥ 6 of 10 on obstacles | Not reached |
| G5 student | Teacher–student gap ≤ 10% success | Not reached |
At iteration 2000: success 0.71, tracking 0.728. With the decoder zeroed, success stayed at 0.99, but tracking fell to 0.027: it stood. That's the paper's ablation, reproduced. But it bounded at every speed.
Both arms avoided the trot decoder. My two-contact projection can't express double support, so every library trot had a flight phase. A four-contact trot (library_v3, 37% with double support) fixed it: forced-trot tracking now fell from 0.47 to 0.15 with ff=0.
Success 0.865, tracking 0.782. With ff=0: 0.891 and 0.023. It handled rough ground and slopes well. Down-stairs fell from 0.93 success on flat rows to 0.12 at levels 7–10. Every one of those steps was a bound.
G3 asks whether the policy picks the gait that suits the situation. It doesn't, yet. At the end of the gait-term anneal, on the paper's obstacle terrain, p(bound) sat at 0.52–0.53 in every speed bin and at 0.46–0.56 on every terrain type. It switched 0.79 times a second, which satisfies the diversity term without meaning anything. Looking for why turned up three bugs of my own:
Watching the final a_scale 0.07 checkpoint terrain by terrain showed a steady trot on rough ground. On stairs, the one real switch to bound came during a descent (the clip below). Stepping stones were messy. On gaps, it walks to the edge, hesitates and falls in, ten times in 8 s, and never jumps. On Oct 2, prompted by the videos, I fed it a fake flat-ground height scan. Success went from 0.684 to 0.682: it had been ignoring the scan. Isaac Lab's default scan sees about 0.8 m ahead, 0.3–0.8 s at these speeds, so reacting blind was cheaper. Two runs started that day with the paper's geometry, a longer scan plus a lidar-like strip to 5 m. Their gate at 3,000 iterations: the fake scan has to hurt.
Everything here is Isaac Lab simulation. Success rates use the paper’s protocol: all obstacle terrains, levels 0–10, commands of 0–2 m/s.
Watching the clips, I suspected the policy wasn’t looking. Replacing its height scan with fake flat ground took success from 0.684 to 0.682: it wasn’t. Isaac Lab’s default scan sees about 0.8 m ahead, 0.3–0.8 s at these speeds, so reacting blind was cheaper. With the paper’s geometry, a 2.3 m scan plus a lidar-like strip out to 5 m, the fake scan finally hurt, most on the hardest rows. The same day I found that Isaac Lab’s “inverted pyramid” is the one the robot climbs, so every earlier “down-stairs” number was really up-stairs.
A per-terrain blind test showed where it looked: descending stairs fell from 0.91 to 0.16 without vision and high steps from 0.93 to 0.39, but stepping stones didn’t change. The legs were brushing obstacles on 2.4 of every 10 control steps, unpenalised, so feeling its way was cheap. Restoring the paper’s leg-collision termination made climbing depend on vision for the first time: 0.91 with it, 0.00 without. I also added the paper’s concurrent estimator and terrain encoder, swapped the scan for a patch matched to what this robot’s D435 and Mid-360 can see (the sensor head), and set a 0.15 m minimum gap, since narrower gaps fell between scan rays.
The new stack looked broken: 0.17 success. It wasn’t. The leg-collision check read the contact sensor’s force history, which still held forces from before the reset, so half of all episodes ended on their first step, in training as well. A three-step grace period fixed it, and the same checkpoint re-scored 0.546.
“Why such small strides?” Every policy since the first walker had shuffled: 0.15 m per cycle at 5–7 Hz, against the library trot’s 0.56 m at 1.7 Hz. Nothing tied the policy’s timing to the library’s rhythm. An air-time reward only nudged it. A contact-schedule reward at the library cadence did it: a clean diagonal trot at 2 Hz, 0.34 m per cycle at 1 m/s, body level, and 0.785 success at 3,000 iterations.
A warm start with commanded gaits got both to work when asked: 0.735 success forced to trot and 0.718 forced to bound. Left to itself, it still never picked bound. I also found that the paper anneals its gait-diversity term over 50,000 epochs; I had used 12,000. And both gaits refused 2.5 m/s: they turned and stepped in place.
Three causes. The library’s trot stopped at 2 m/s, with only 476 bounds above 2.5. My generator checked torque and joint speed as separate limits, but the A1’s motor delivers less torque the faster it spins, about 33.5 × (1 − speed/21) N·m; by that test 54–86% of the fast bounds were infeasible. And the gait clock was fixed. Library v6 uses the exact motor curve: about 189,000 trajectories, 87,215 trot and 101,835 bound. Nothing is feasible above 3 m/s, and 2.5–3 m/s needs a low swing. The VAE fits it with R² 0.963.
A run with commanded gaits tracked 0.91–0.93 of commands from 2.1 to 3.2 m/s, reaching 2.43 m/s at the top band, where earlier runs managed 0.05. But its obstacle success fell to about 0.48. A warm start recovered it only partly, to 0.55, while keeping the speed.
A 50,000-iteration run with the policy choosing its own gait and the paper’s full anneal reached 0.835 success, the best yet. On the hardest stair rows it scores 0.78–0.84, high steps 0.93, hurdles 0.81, rough ground 0.98. It gave up three things. It trots 93% of the time at every speed and terrain. Its speed ceiling came back: at a 2.6–3.2 m/s command on flat ground it does 1.15 m/s, probably having learned that a high command on hard terrain means slow down. And stepping stones fail above level 4. The velocity estimator doesn’t matter: feeding it ground truth instead moves success from 0.841 to 0.846.
Its gap “successes” are it standing at the edge. A gap-only run showed why: a trotting foot drops into a 0.15 m slot about half the time, the calf brushes the wall, and the episode ends. Every early attempt to cross died, so it learned that a gap means stop. Two runs are now training with leg contact on gap tiles as a penalty instead of a termination, plus the paper’s command coupling: slower on high stairs, no sideways or turning commands above 2 m/s.
| Departure | What forced it |
|---|---|
| Speed to 2 m/s (bound to 3), slower trot timing | The A1's 1.050 s leg pendulum; joint velocity rejected nearly half of all samples |
| Library filtered to A1 limits, 25% torque in reserve | Stored torques alone hit the ceiling in 92% of fast replays |
| Load split in place of initial pitch | The paper's stated unknowns gave a singular solve here |
| Four-contact trot | RL avoided a trot decoder that only knew running trots |
| kp 37 / kd 0.9 (paper 80 / 2), a_scale 0.07–0.1 | The A1's deployable gains; the auxiliary channel's authority has to stay relative to decoder torque |
| legged_gym-style reward clip, penalty ramp | Without them early life is net-negative and PPO dies or stands |
| Gait anneal compressed to the run length | 50,000 epochs is beyond my budget |
| Foot-collision by contact direction | Height-based versions punished standing on obstacles |
These are the reports I wrote during the work, unedited, as they were at the time. Several were later corrected or retracted. The tags say which.
Library trajectories on the articulated A1, and the torque-sign bug. Its survival rates turned out to measure open-loop playback, not library quality.
The 35% tracking win. Retracted that morning: the auxiliary channel was supplying 94% of the torque.
The first torque-split audit on the parkour stack. This run later turned out to have trained through the reward blackout, so its numbers are void.
Interactive: all 47,174 lib_v2 trajectories by speed, period, duty and height, 1,083 of them playable, with contact, torque and force traces.
The plain parkour teacher at terrain 2.103, the reference every phase-one library arm was measured against.
The correction: 44.5% gait mixing at 0.097 m/s, and how a reward term paid the policy not to walk.
Three real bugs, seven changes at once, and the 0.537 m/s walker. Its video was replaced after the face-plant mix-up, and the page says so.
Exploration versus incentive: p(bound) 0.010 against 0.568, with both final videos frame-checked.
Open-loop authority, phase structure, the trot bank. Its morning "architecture verdict" was confounded by a 15 cm spawn drop, found on Sep 29.